A comprehensive GeoAI review: Progress, Challenges and Outlooks

Anasse Boutayeb * Iyad Lahsen-cherif * Ahmed El Khadimi *

December 17, 2024

Abstract- In recent years, Geospatial Artificial Intelligence (GeoAI) has gained traction in the most relevant research works and in-
dustrial applications, while also becoming involved in various fields of use. This paper offers a comprehensive review of GeoAI as
a synergistic concept applying Artificial Intelligence (AI) methods and models to geospatial data. A preliminary study is carried out,
identifying the methodology of the work, the research motivations, the issues and the directions to be tracked, followed by exploring
how GeoAI can be used in various interesting fields of application, such as precision agriculture, environmental monitoring, disaster
management and urban planning. Next, a statistical and semantic analysis is carried out, followed by a clear and precise presenta-
tion of the challenges facing GeoAI. Then, a concrete exploration of the future prospects is provided, based on several informations
gathered during the census. To sum up, this paper provides a complete overview of the correlation between AI and the geospatial
arXiv:2412.11643v1 [cs.AI] 16 Dec 2024

domain, while mentioning the researches conducted in this context, and emphasizing the close relationship linking GeoAI with other
advanced concepts such as geographic information systems (GIS) and large-scale geospatial data, known as big geodata. This will
enable researchers and scientific community to assess the state of progress in this promising field, and will help other interested parties
to gain a better understanding of the issues involved.

Keywords : Artificial Intelligence (AI), GeoAI, geospatial data, GIS, big geodata.

1 Introduction digital transformation, and to ensure sustainability and global de-

velopment.
Artificial intelligence (AI) has become one of the most influential More specifically, AI is a key tool allowing the management and
technological advances in human history [1], involving radical exploitation of geospatial data, paving the way for a new era of
metamorphoses in numerous active sectors, such as healthcare, analysis and interpretation, particularly when it comes to massive
finance, commerce, transport, etc. AI relies on several algorithms and complex data, or when it comes to sophisticated tasks. As a
and methods to emulate human cognition, enabling data analysis result, AI can perform these tasks while guaranteeing high accu-
in any quantity or complexity, achieving advanced automation of racy, including the processing of remote sensing images to pro-
the most complicated processes, and helping to make the most duce land use maps [16], the segmentation of agricultural crops
crucial decisions. It can also perform various tasks with unprece- [17], the detection of cars on road networks [18], the analysis of
dented precision. For example, AI facilitates the exploitation of changes in spatio-temporal time series [19], and the use of pre-
medical images while enabling the automatic and rapid detection cipitation to make weather forecasts [20]. Additionally, powerful
of tumours; it also facilitates the analysis of the most complex AI models are designed to exploit temperature and vegetation in-
genetic data [2]. On the other hand, AI algorithms optimise pro- formations in order to quantify the degree of drought over a vast
duction and exploitation of renewable energies, predict consumer geographical area, massive trajectory data can also be exploited
demand and the optimum location for a large surface area [3]. to model high-performance route calculation algorithms. AI can
In addition, AI can make an effective contribution to improving also be used to predict the weather of floods by analysing remote
teaching methods, by adapting the pace and content of courses to sensing images and rainfall level measurements [21]. It should be
students’ levels, while analysing consumer purchasing behaviour noted that this multitude of applications of artificial intelligence
[4], improving sales techniques and boosting yields [5]. In short, in relation to geospatiality has given rise to a new concept, namely
AI is currently a promising vector for efficiency and development, GeoAI, which is carefully explored in this paper, indicating the
guaranteeing exceptional performance. issues involved, the level of progress achieved and futurs trends.
Furthermore, geospatial information has a particular impact on The remainder of this paper is organized as follows: section 2
people’s daily lives, making it easy to navigate in urban and rural presents the issues at stake in the subject, the Research Questions
areas, locating schools, parking lots, stores, hospitals and medical (RQs) to be answered and the research directions that have fol-
centers with great precision, and finding directions to these places lowed, section 3 introduces the theme of artificial intelligence,
very quickly. In addition, geospatial technologies are at the heart covering the main concepts and models, as well as new emerg-
of cutting-edge fields such as agriculture [6, 7], environment [8, ing approaches, section 4 develops the research directions defined
9], defense [10], urban planning [11, 12, 13] and healthcare [14, above, leaving the final section to provide exhaustive answers to
15], while offering a new dimension of analysis and reflection, by the research questions posed, and to introduce the related chal-
combining information from various acquisition systems, specif- lenges and the future prospects, presented as results of the explo-
ically satellites, drones, GPS receivers, etc. This enables a high- ration carried out.
level community of researchers and industrialists to address to-
day’s global challenges in every field, to participate in the era of

* Institut National des Postes et Télécommunications (INPT)

2 Issues, key questions and research di- i.e. each sub-domain to be developed is completely distinct, in
terms of research objects, from the others. Furthermore, and to
rections situate the aim of this work in relation to previous works, a great
deal of research is carried out on this subject, as shown in Table 1,
2.1 History and definitions while touching on different sub-branches of geospatial field, but
Looking at history of the synergy between AI and geospatial area, without giving a detailed overview, nor focusing on a wide range
it is noted that this coupling began as early as the 80s and 90s, of application areas.
through the works of Smith [22], Estes et al. [23], Couclelis [24], The problem at hand, is to provide a comprehensive review of
Openshaw and Openshaw [25]. A major milestone in the history the GeoAI’s state of the art in terms of application domains, dif-
of GeoAI is the ”Association for Computing Machinery (ACM) fering from previous work both in terms of the completeness of
SIGSPATIAL conference” in 2017, where a specific workshop is the inventory and in terms of research objectives.
dedicated to the subject, defining GeoAI explicitly as the coupling To address this issue, a scoping and design phase is carried out,
of Artificial Intelligence techniques to geographic information or which consists of choosing the working methodology and clearly
geospatial data [26, 27]. Moreover, several definitions of GeoAI identifying the preliminary issues. Then, a definition of research
have been provided. According to Gao [28], GeoAI can be re- questions is fundamental to orientate the census to be carried out.
garded as : Next, a determination of research axes is worked out to identify
key elements of the concerned investigation, using an advanced

“ a branch of AI whose objective is to imple-

ment intelligent computer programs aiming to
imitate human perception, particularly with
keyword search on Scopus database [33] and Google Scholar en-
gine [34], on the one hand, and refining the choice of fields of
application to be discussed, on the other. As illustrated in Figure
2, the papers are selected on the basis of a set of criterias:
regard to spatial reasoning and geographical
dynamics. • Innovative contribution: given that the final objective of this

Li and Hsu [29] also define GeoAI as :

” work is to explore the level of progress achieved.
• Diversification of selection themes: the aim is to cover all areas
in question without duplicating the subjects tackled.

“ a new field of interdisciplinary research that

uses geospatial Big Data while leveraging and
developing AI for location-based analysis.
• Quality of papers: those mainly published by renowned pub-
lishers such as Elsevier [35], the Institute of Electrical and Elec-
tronics Engineers (IEEE) [36], Springer [37] and Multidisci-

”
Furthermore, emphasizing the applicative aspect of the geospatial
plinary Digital Publishing Institute (MDPI) [38] are selected.

2.3 Research questions and directions

field, Kausica et al. [30] assert that :
A crucial step in this research is to first formulate the research

“ GeoAI combines AI algorithms with geospa-

tial methods to extract the most significant in-
formation, with the purpose of performing so-
questions in order to clarify the objectives to be achieved and to
guide the work to be carried out. Three questions are important
to ask:

phisticated tasks in the same field. • RQ1: How can GeoAI methods contribute to effective and ac-

2.2 Related topics, problematic and census ap-

” curate results in various fields of application?
• RQ2: What are the challenges currently facing GeoAI?
• RQ3: What are the future prospects for GeoAI? More specifi-
proach cally, what are the promising themes to be explored in this con-
As presented above, GeoAI highlights a fundamental aspect that text?
will be exploited in this paper, namely the synergy between AI
Throughout the rest of this paper, relevant research focusing on
and its most advanced techniques, on the one hand, and geospa-
the application of creative AI methods to geospatial domain are
tial data, on the other. At a deeper level, the importance of linking
identified. To this end, the following research directions are cho-
GeoAI to geographic information systems (GIS) should be noted,
sen for exploration, according to their growing importance and
as this enables geospatial data to be catalogued for appropriate
their presence in the most relevant research works:
manipulation, analysis and presentation [31]. Figure 1 illustrates
the contribution of GIS as an ideal collaborator with AI, due to its • Precision agriculture;
efficiency and standardization.
Another fundamental aspect to highlight is the use of big geodata, • Urban planning, logistics and transportation;
a concept referring to large-scale geospatial data, which empha-
sises the use of AI models to manage, process and use this type of • Environmental management;
data, taking into account its variety of sources, natures and scales • Water resources management and precipitation forecasting;
[32].
To address the subject of GeoAI comprehensively, it is essential • Natural disaster monitoring;
to emphasize the importance of citing the various applications
of GeoAI, including urban planning, logistics and transportation, • Healthcare.
agriculture, water resource management, and many others. In
fact, the importance of listing GeoAI research works by applica-
tion domains allows to follow a semantically consistent approach,
2
 Figure 1: GeoAI and GIS.

The definition of these axes is intended to give an overview of Furthermore, LeCun et al. [41] presents Deep Learning (DL) as :
the state of the art of GeoAI, not only in terms of the multitude
of functionalities it offers, but also in terms of the diversity of its
fields of application and its contribution to geospatial area, as well
as the relevance of messages and services it provides.
Before getting to the heart of the chosen research axes, it is needed
“ a branch of Machine Learning using multi-
layer architectures to learn multiple represen-
tations of data, it is noted that deep learning
to present the basic principles of artificial intelligence, its def- uses the back-propagation technique to ad-
initions, methods, approaches and evaluation metrics, which is just the internal parameters of the model in
conducted in the following subsections (from 3.1 to 3.4). Sub- order to recalculate these representations at
section 3.5 is devoted to the presentation of the commonly used each epoch.
datasets, afterwards, subsection 3.6 is dedicated to the hardwares
used, its categories, architectures, specifications, etc. Section 4
then develops the above-mentioned research directions, with each
”
In this section, a comprehensive overview of all commonly used
subsection listing the various works presenting GeoAI models ac- AI methods and models is given, along with the main concepts
cording to the relevant field of application, respecting the selec- put forward in the context in question.
tion criteria indicated in 2.2 in order to provide a comprehensive
answer to RQ1. Finally, section 5 is dedicated to the answers to 3.1 Machine Learning (ML) models
RQ2 and RQ3, giving a detailed presentation of the challenges
and the perspectives involved. In order to carry out the various tasks commonly performed by
AI, such as regression where the values of scalars are predicted,
classification where a label is assigned to a specific piece of data,
3 Artificial Intelligence : definitions and or clustering where a partitioning of data is proposed, a multitude
of machine learning models are presented in the literature. In this
models subsection, different models are explored depending on the nature
of the data used. In fact, two types of learning are distinguished:
According to Xu et al. [39] Artificial Intelligence (AI) corre-
supervised learning, when the training data are labeled, and unsu-
sponds to :
pervised learning, in the opposite case [42].

“ a discipline referring to the simulation of hu-

man intelligence by machines, with the aim
of imitating human behavior in different situ-
3.1.1 Supervised learning
In this context, the data used for training is labeled, i.e. each data
record is associated with a specific value. This value is then em-
ations.

”
Moreover, Santosh et al. [40] define Machine Learning (ML) as :
ployed during the learning phases as a reference value, used to
represent the links between inputs and outputs [43]. In this sub-
section, the most notable regression and supervised classification
models are explored.

“ a set of techniques for approximating a func-

tion that maps an input space to an out-
put space, while extracting meaningful, non-
Linear Regression: is an approach for modelling the relation-
ship between a scalar dependent variable and explanatory vari-
ables, the case of a single explanatory variable being called sim-
redundant information from data samples. ple linear regression [44], the relative equation is as follows:

” ŷ = β0 + β1 x1 + β2 x2 + · · · + βn xn + ε, (1)

3
 Table 1: Previous research on GeoAI and GIS-AI synergy

References Paper title Journal/Conference Object

[45] GeoAI at ACM SIGSPATIAL: progress, challenges, SIGSPATIAL Special (2019) Explore the workshops from the 2017 SIGSPATIAL conference while covering several application areas,
and future directions including transport and public health.
[46] A five-year milestone: reflections on advances and Annals of GIS (2024) Determining the level of progress and constraints of the GeoAI five years on after the first American Associa-
limitations in GeoAI research tion of Geographers (AAG) symposium.
[47] GeoAI – Accelerating a Virtuous Cycle between AI IC3-2021: Proceedings of the 2021 Thirteenth List the possibilities offered by AI in the geospatial field, identifying the areas of application of GeoAI, such
and Geo International Conference on Contemporary Com- as geocoding and change detection.
puting (2021)
[48] GeoAI: a review of artificial intelligence approaches Geoscientific Instrumentation, Methods and Data Use the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) standard to system-
for the interpretation of complex geomatics data Systems (2022) atically explore GeoAI techniques according to the type of geospatial data used, while developing a statistical
analysis.
[49] On the Opportunities and Challenges of Foundation ACM Transactions on Spatial Algorithms and Explore the prospects of possible GeoAI foundation models for different categories of geospatial data.
Models for GeoAI (Vision Paper) Systems (2024)
[50] GeoAI and its implications International Encyclopedia of Geography: Peo- Determine the degree of progress of GeoAI in many fields of application and explore its integration into the
ple, the Earth, Environment and Technology conceptualisation of geospatial methods.
(2016)
[51] GeoAI: Where machine learning and big data con- Journal of Spatial Information Science (2020) Exploring the contribution of GeoAI to the processing and analysis of big geodata, while identifying its chal-
verge in GIScience lenges and future promises.
[52] Exemplification on Potential Applications and Sce- 2023 Asia-Europe Conference on Electronics, Determine the current advances in GeoAI research, while listing its latest practices in relation to geospatial
narios for GeoAI Data Processing and Informatics (2023) methods.
[53] GeoAI: spatially explicit artificial intelligence tech- International Journal of Geographical Informa- List the various applications of AI in geospatial data processing, as well as the issues involved.
niques for geographic knowledge discovery and be- tion Science (2020)

4
yond
[54] A review of recent researches and reflections on Geomatics and Information Science of Wuhan Present the relationship between GeoAI and the advancement of geospatial research, exploring challenges and
geospatial artificial intelligence University (2020) outlooks.
[55] Special issue on geospatial artificial intelligence Geoinformatica (2023) Explore all the papers presented at the 4th edition of ACM SIGSPATIAL related to the subject of GeoAI.
[56] Applications of Artificial Intelligence and Machine Emerging Trends, Techniques, and Applications Discover the advances in the application of machine learning (ML) techniques, by presenting the contribution
Learning in Geospatial Data in Geospatial Data Science (2023) of GeoAI to the extraction and generation of geospatial information.
[57] GeoAI and deep learning The International Encyclopedia of Geography Present very quickly the close relationship between deep learning and the geospatial domain.
(2021)
[58] Philosophical Foundations of GeoAI: Exploring Sus- Handbook of Geospatial Artificial Intelligence Discover the philosophical foundations of the GeoAI while highlighting fundamental issues such as the lack
tainability, Diversity, and Bias in GeoAI and Spatial (2023) of ethical neutrality of the GeoAI and its sustainability.
Data Science
[59] Artificial Intelligence for Multisource Geospatial In- ISPRS International Journal of Geo-Information Exploring the application of various artificial intelligence (AI) methods to geographic data from a multitude
formation (2022) of sources.
[60] Geospatial AI in Earth Observation, Remote Sensing, Applied Sciences (2023) Determine the progress of GeoAI in relation to GIS and remote sensing.
and GIScience
[61] A Brief Review of Recent Developments in the Inte- Geomatics and Environmental Engineering Provide a partial overview summarising the synergy between Deep Learning (DL) and GIS, while discovering
gration of Deep Learning with GIS (2022) its impact on several fields of application such as hydrology and natural disasters.
[62] Merging GIS and Machine Learning Techniques: A Journal of Geoscience and Environment Protec- Discover several research projects combining Machine Learning (ML) methods and GIS, as well as the appli-
Paper Review tion (2022) cations of this synergy, such as health and erosion modelling.
 • x1 , x2 . . . xn are the independent variables used,
• β0 , β1 . . . β1 are the coefficients to be adjusted.

Decision trees: is a model used for regression and classification

tasks, the principle behind which is to divide the training data into
sub-batches for each node of the tree, with a set of decision cri-
teria defining the depth of the tree until the final predictions are
reached [64].

Random Forest (RF): is a supervised model combining several

decision trees to perform predictive tasks such as classification,
regression, etc. Introduced by Breiman [65], the principle of Ran-
dom Forest consists in training each tree from a batch of data by
random selection, the final result is then obtained from an average
of results from all the trees, Figure 4 presents the principle of this
model.

Figure 2: Review workflow.

where:

• y is the variable to be predicted,

• ŷ is the value of y predicted by the model,
• x1 , x2 . . . xn are the independent variables used,
• β0 , β1 . . . βn are the coefficients of this regression,
Figure 4: Random Forest illustration ([66]).
• ε is the error committed by the regression.
The aim of this model is to learn the regression coefficients Support Vector Machine (SVM): proposed by Cortes and
β0 , β1 , . . . , βn by minimizing the value ∑m 2
i=1 (yi − ŷi ) , where m Vapnik [67], it is a model used for classification. The mathe-
is the number of data points. matical principle is to search for the hyperplane (n-dimensional
In the case of a simple Linear Regression, Figure 3 shows the line generalization of a plane) that best maximizes the distance be-
that best fits a data point, β0 is the y-intercept and β1 is the slope tween the hyperplane and the various data classes.
of the blue line. Figure 5 illustrates SVM model for a binary classification prob-
lem, the optimal hyperplane is constructed to separate the positive
classes from the negative ones as best as possible, while maximiz-
ing the margin between its classes, two supporting hyperplanes
(dotted lines) define the positive and negative boundaries.
For the multi-class problem, several algorithms are used to trans-
form it into several binary classification problems, such as One-
vs-Rest and One-vs-One [68].

Figure 3: Example of linear regression.

Logistic Regression: proposed by Cox [63], logistic regression

calculates the probability of a variable belonging to a given class
by optimizing the involved coefficients. The associated probabil-
ity function is a sigmoid function:
1
P(y = 1|x) = , (2)
1 + e−(β0 +β1 x1 +β2 x2 +···+βn xn )

where: Figure 5: SVM ([69]).

• y is the binary variable to be predicted (class membership),

5
K-Nearest Neighbors (KNN): is a model used for classifica- The K value is a key parameter in K-means algorithm, as it refers
tion and regression, as shown in Figure 6, KNN consists in iden- to the number of clusters, and therefore the clustering structure
tifying, for a given measure, the K closest points in terms of dis- implemented, thus, it directly affects the model performance in
tance [70]. The value assigned to the candidate point is deter- terms of prediction quality and convergence.
mined by the most frequent label of the K points, in the case of
classification, and by averaging the K neighboring values, in the Density-Based Spatial Clustering of Applications with Noise
case of regression. (DBSCAN) is a clustering model proposed by Ester et al. [77],
enabling a set of data points to be grouped together on the basis
of their density. The principle of this algorithm consists in using
a central point while defining a neighborhood radius, and a mini-
mum number of points to be grouped together in each cluster. The
strength of DBSCAN lies in the fact that it enables measurement
points of various shapes and sizes to be grouped together, and does
not require the number of resulting groupings to be specified.

3.2 Deep Learning models

These are model architectures mimicking the functioning of the
human brain. Deep Learning (DL) uses Deep Neural Networks
(DNN) as a basic approach, highlighting two key concepts: gra-
dient descent to calculate the weights of each neuron, and back-
Figure 6: KNN mechanism.
propagation to calculate the gradients of the cost function by in-
volving the propagation of error backwards [78]. In this section,
Gradient boosting: is a supervised technique used for classi- the main Deep Learning models are listed.
fication and regression, which involves combining the prediction
results of several decision trees. It is noted that Gradient boosting 3.2.1 Multi Layer Perceptron (MLP)
refines the prediction results of subsequent trees on the basis of
the previous ones. Proposed by Rosenblatt [79], MLP represents the most elemen-
Gradient boosting uses ensemble learning, the principle of which tary Deep Neural Network architecture. The first layer transmits
is to use several ML algorithms to obtain, via several voting mech- input data representation to the hidden layers, where each neuron
anisms, results that are better than those obtained by each of the is connected to all neurons in the next layer. This transmission
methods alone. Several methods are used in ensemble learning, of information between layers is based on non-linear activation
such as Bagging [71], Boosting [72] and Stacking [73]. functions applied to each neuron, enabling the model to better un-
More specifically, Extreme Gradient Boosting (XGBoost) is an derstand relevant dependencies between data. Other probabilistic
optimized implementation of the Gradient Boosting algorithm. activation functions are applied to output layer in the case of clas-
Introduced by Chen and Guestrin [74], XGBoost stands out for sification, or a linear activation function in the case of regression,
its computational efficiency and resource management. in order to achieve meaningful results [80].
In spite of its simplicity, MLP is used in particular for tasks such
3.1.2 Unsupervised models as games, text and voice recognition.

As stated by Barlow and Horace [75], unsupervised learning al- 3.2.2 Convolutional Neural Networks (CNN)
gorithms allows to learn the characteristics and representations
of data without having any information about data labels. This Convolutional Neural Network is a specific Deep Neural Network
seems very practical, given that, in many cases, the data are not (DNN). As shown in Figure 7, it is broken down into the following
annotated. In the following, unsupervised ML models are listed. elements, convolutional layers for Feature Extraction (FE), pool-
ing layers for dimension reduction and Fully Connected layers
K-means is a clustering method grouping a set of data points (FC) for result classification. It is noted that several researchers
into K groups. Each measurement is taken and assigned to the have introduced the concept of CNN, such as Zhang et al. [81] and
nearest group, by measuring the distance between the point in LeCun et al. [82], but the first work to implement a base of CNN is
question and the centroid of each group, then, this centroid is Waibel et al, who in 1987 proposed a Time-Delay Neural Network
recalculated taking into account the new point added. This op- (TDNN) [83], which can be considered as a one-dimensional con-
eration is repeated until the shape of groups becomes immutable volutional neural network. CNN have demonstrated notable effi-
[76]. Mathematically, groups are formed by minimizing the sum ciency for several computer vision tasks, including classification,
of the squared distances between the measurements and the cen- segmentation and object detection. In this context, a multitude
troids of the groups, i.e: of models have been developed, such as VGG [84], U-Net [85],
ResNet [86] and YOLO [87].
K
∑ ∑ ∥x − µi ∥22 , (3)
3.2.3 Recurrent Neural Networks (RNN)
i=1 x∈Si
RNN is a class of deep neural networks designed specifically to
where: process sequential data including text, voice, time series, etc. In-
• x is the measurement in question, troduced in numerous works such as those by Hopefield [88] and
Elman [89], The special feature of the RNN is the recursiveness
• µi is the centroid of the cluster, of its connections, enabling this type of network to remember pre-
vious entries.
• ∥.∥2 is the L2 norm.
6
 Figure 7: CNN architecture.

Despite the flexibility and adaptability offered by RNN, this archi- that distinguishes between ”false” and ”real” data, enabling the
tecture faces several problems, particularly when processing long generator to improve accuracy of the data it produces. Figure 9
sequences, specifically the problems of vanishing and exploding describes this process.
gradient [90], as well as the problem of retaining long-term de- Several applications use the basic GAN architecture, such as art
pendencies [91], and the drain on computational resources due to [100], image enhancement [101, 102], Video synthesis [103] and
long-term training [92]. These problems are overcome using the image translation [104].
LSTM model [93], a special RNN architecture made up of several
blocks. Each block deals with a given stage of the sequence using
four key elements:
• Cell state, transmitting dependencies throughout the se-
quence.

• Forget state, this component determines the information to

be “forgotten” from cell state.
• Input gate, deciding which information to add to cell state.
• Output gate, sdetermining the information to be used for
the final output.
Figure 9: GAN workflow.

3.2.4 Auto-Encoders (AE)

3.3 Advanced notions of Artificial Intelligence
It is a particular neural network for unsupervised generative tasks.
Presented by Hinton et al. [94], the principle is to pass the data In addition to Artificial Intelligence architectures, it is worth men-
representation into a latent space in order to reduce dimensual- tioning a number of advanced concepts linked to this subject.
ity, and then reconstruct the data by transposed convolution. As These concepts refer to mainly “theoretical” approaches, allow-
illustrated in Figure 8, the encoder compresses the data while ex- ing traditional algorithms and methods to make unprecedented ad-
tracting the essential features captured, whereas the decoder aims vances. Indeed, reinforcement, federated, ensemble, multimodal
to reconstruct the input structure through these features. and transfer learning are presented, along with attention mecha-
Auto-encoders are used for several tasks such as image compres- nisms.
sion [95], anomaly detection [96] and denoising [97].
3.3.1 Reinforcement learning
Reinforcement learning is an ML technique in which the agent
improves iteratively its behavior on the basis of rewards or penal-
ties considered as the results of its actions, representing the
agent’s interaction with a given environment, This environment,
modeled by a set of states, stands for the external setting to which
the agent tries to evolve its future state. To do this, the agent tries
to improve its decision making during the learning process[105].
Numerous variants of reinforcement learning are proposed,
such as State-Action-Reward-State-Action (SARSA) [106], Q-
Learning [107] and Proximal Policy optimization [108].
Figure 8: AE architecture.

3.3.2 Federated learning

3.2.5 Generative Adversarial Networks (GAN)
Introduced by Mcmahan et al. [109], this innovative approach en-
As their name suggests, Generative Adversarial Networks are ables multiple entities (physical or logical), to contribute to the
used to generate data that resembles real ones as closely as pos- training of a model without exchanging directly users data. Each
sible. Proposed by Goodfellow et al. [98], a GAN consists of entity produces its own updated widgets based on this training, a
two networks, a generator creating data from a latent space, of- central server combines then these updates to improve the model
ten a Gaussian or uniform distribution [99], and a descriminator
7
under consideration. Figure 10 summarizes this process in three fore calculating attention scores for each patch, followed by nor-
steps. malization using a probabilistic function for gradient stabilization
and scaling, then, weighted vector calculation and linear aggrega-
tion are implemented to combine the relevant extracted represen-
tations.

Figure 10: Federated learning workflow.

One should note that federated learning architecture offers a mul-

Figure 11: Attention mechanisms
titude of remarkable attributes, including reduced communication
.
costs, task scalability, security, confidentiality and data decentral-
ization, but faces a multitude of problems such as data imbalance
Many creative models are based on these mechanisms, for exam-
and system failures [110].
ple:

3.3.3 Multimodal learning

Transformers: introduced by Vaswani et al. [119], transform-
According to Ngiam et al. [111], multimodal learning involves ers are architechtures used mainly in Natural Language Process-
the linking of information from multiple sources, for example, ing (NLP), but adapted to multiple applications such as computer
several categories of data including visual data, sound and text vision, via Vision transformers (ViTs) [120].
can be used to perform a specific task, such as recognising the ex- As indicated in Figure 12, a transformer consists of an encoder
act representations of words or analysing feelings [112]. processing the inputs and a decoder generating the output. It is
Several approaches fit into the multimodal learning, including the noted that each of these two components is made up of a single
early fusion, where multimodality is used from the start of model layer of feed forward neural networks, as well as residual addition
training, by combining features extracted from different sources and normalization to preserve the information transmitted by the
of trained data, and the late fusion, where each modality is trained model, along with positional encoding to manage sequence order.
separately, the results of each prediction are then weighted to- Each encoding-decoding layer is reinforced by a self-attention
wards the end of the process [113], as well as the shared repre- mechanism in order to extract dependencies between sequence
sentation learning, where all the modalities are projected into a components, including image patches, tokens, etc. In the attention
single latent space. The aim of the latter approach is to capture the mechanism in question, three vectors are calculated to compute at-
relationships between these different modalities in order to align tention scores, the Query (Q) representing the searched element,
and exploit them optimally [114]. the Key (K) reflecting the input element, used by the query, and
the Value (V) vector referring to the actual value related to the
3.3.4 Transfer learning processed component of sequence.

To overcome the problem of lack of data for a specific task, trans-

fer learning exploits data relating to a source task or obtained
in domains related to the main task [115, 116]. In practical
terms, this involves adjusting the weights of an initial model, of-
ten trained on a large scale dataset, to a much smaller one.
According to Weiss et al. [117], transfer learning improves learn-
ing while improving performance with respect to the configura-
tion offered, by making better use of the training data, and by
enormously reducing the cost of calculating weights. Despite this,
transfer learning faces several challenges, in particular the nega-
tive transfer, referring to the case where the target domain is too
different from the source one, as well as the non-availability of
labelled data, whether for the source or target domain, a problem
that is proving difficult to overcome.

3.3.5 Attention mechanisms

It is an advanced technique allowing a model to concentrate on
the essential parts of data. The contribution of this concept is to
be able to focus on relevant data by imitating the way humans
pay attention to the interesting elements around them [118]. Fig- Figure 12: Transformers architecture from [119]
ure 11 describes an example of attention mechanism with respect .
to the constituents of a sequence, for instance, in this schema, re-
motely sensed image patches. Vector embedding is performed be-

8
Large Language Models (LLMs): is a family of NLP models 3.4.4 Determination coefficient R2
generating coherent, meaningful text from deep neural networks.
R2 is a statistical measure, introduced by Wright [126], to mea-
The associated architecture is characterized by the use of attention
sure the degree to which a regression model truly describes the
mechanisms and a very large-scale text data. Several LLM models
relationship between its variables, by calculating the proportion
have revolutionized the field of NLP, such as Text-to-Text Transfer
of variance of the dependent variable from independent variables
Transformer (T5) [121] and Generative Pre-trained Transformer
in the involved model. Mathematically , this measure is calculated
(GPT) [122].
as follows:
∑n (yi − ŷi )2 ,
R2 = 1 − i=1 (8)
3.4 Evaluation metrics ∑ni=1 (yi − ȳ)2
A variety of metrics are employed to evaluate the accuracy of
where, for an observation i:
models, depending on the nature of concerned tasks. In this sub-
section, the metrics mentioned are presented in order to give the • yi is the real value of the dependent variable,
reader a clear and concise overview.
• ŷi is the value predicted by the model,
3.4.1 Mean Absolute Error (MAE) • ȳ is the average of the real values over the n observations of the
This is a simple metric for calculating the accuracy of a regression dependent variable.
model, by measuring the magnitude of model’s prediction errors
[123]: 3.4.5 Confusion matrix
1 n This is a visualization of the predictions of a classification model
MAE = ∑ |yi − ŷi |, (4)
n i=1 with respect to ground truth values. It has the following form:
where: Given class Predicted: Negative Predicted: Positive
Ground truth: Positive False Negative True Positive
• n is the number of observations, Ground truth: Negative True Negative False Positive

Confusion matrix gives a detailed overview of the predictions ob-

• for an observation i:
tained, while clearly identifying the best predicted classes and the
– yi and ŷi are respectively the actual and the predicted errors made [127].
value.
– |.| is the absolute value. 3.4.6 Precision
Precision is an indicator measuring the degree of identification of
3.4.2 Mean Square Error (MSE) positive predictions for a classification model [128]. In the case
of binary classification, it is obtained as follows:
Used also to measure the quality of a regression model [124], the
formula involved is as follows: TP ,
Precision = (9)
1 n T P + FP
MSE = ∑ (yi − ŷi )2 , (5)
n i=1 where TP represents the number of true positives, i.e. cases cor-
rectly identified as positive, and FP is the number of false pos-
Root Mean Square Error (RMSE) is a metric estimating the mag- itives, i.e. cases incorrectly identified as positive. In the case of
nitude of prediction errors for a given model [124]. The associated multiple classes, precision is calculated by averaging the precision
formula is: √ obtained for each class using arithmetic or weighted average.
RMSE = MSE. (6)
3.4.7 Recall
3.4.3 Mean Average Precision Error (MAPE)
Recall is another measure quantifying the level at which a classi-
MAPE is a metric used mainly in regression tasks in order to cal- fication model identifies positive classes from the set of positive
culate the predictive accuracy of a given value [125]. This metric instances in measurement sample [128], for binary classification:
gives a normalized measure of prediction errors and is calculated
as follows: TP ,
Recall = (10)
n
T P + FN
100 yi − ŷi
MAPE = ∑ . (7) where FN is the number of false negatives, i.e. cases incorrectly
n i=1 yi
identified as negative.
Comparing the metrics presented so far, it is clear that MAE of- Like precision, recall in generic cases is deduced by averaging the
fers a more direct measure by giving equal weight to all errors, recalls of each concerned class.
that MSE is more sensitive to errors than MAE and MAPE, since
it amplifies the errors measured. Additionally, the RMSE is easier 3.4.8 Accuracy
to interpret than MSE, since it is expressed in the same unit of
observations, while MAPE allows relative error to be measured It is a performance metric, defined as the ratio between the cor-
in percentage terms, offering the possibility of evaluating the effi- rect predictions of a classification model and the total number of
ciency of models driven by data at different scales. predictions, for the case of binary classification:
TP+TN .
Accuracy = (11)
T P + T N + FP + FN

9
3.4.9 F1-score predict never-before-seen scenarios. In addition, a data set must
be well labeled, in the case of supervised learning. For example, a
This metric evaluates the ability of a classification model to ef-
dataset of remote sensing images used for aircraft detection needs
fectively predict positive individuals, by making a compromise
to be able to cope with a multitude of configurations, i.e. hidden
between precision and recall [128]. The combination of precision
aircraft, different markings, different colors, etc.
and recall is presented in the form of a harmonic mean:
Given the specific nature of geospatial data, particularly in terms
Precision × Recall . of data volume and diversity of formats available, standardizing
F1-score = 2 × (12) and adjusting these data is essential for producing accurate GeoAI
Recall + Precision
models. Various geospatial datasets are available, allowing to
F1-score is a metric giving a balanced evaluation of model’s per- train and validate models on one hand, and to compare the per-
formance, useful when there is an imbalance between positive and formance of different AI models on the other.
negative classes. In this subsection, different categories of geospatial data are ex-
plored and the main existing geospatial datasets are examined,
3.4.10 Jaccard index whether datasets proposed as part of challenges, research or
datasets submitted as part of space programs and therefore fed
Jaccard’s index, often known as Intersection over Union (IoU), as and when new acquisitions are made. It is noted that, in many
is an evaluation metric used in segmentation and detection tasks. cases, geospatial datasets are not defined by the number of in-
It relates to the ratio between the intersection and the union of stances, but rather by their geographic coverage and spatial reso-
predictions and the ground truth [129]: lution.
Area of Intersection .
IoU = (13) 3.5.1 Sattelites images datasets
Area of Union
Earth observation sattelites provide various users with image data
3.4.11 mean Average Precision (mAP) at different resolution scales, ranging from low resolution (greater
It is an evaluation metric often used for detection tasks. It cor- than 100 meters) or medium resolution (typically between 10 and
responds to the average value of mean accuracies of each class 100 meters) to very high resolution (less than 1 meter) [132].
[130]: These images, acquired using a variety of techniques, whether
1 N passive cameras such as optical or thermal, or active sensors such
mAP = ∑ APi . (14) as radar, are used to capture terrestrial information and exploit it
N i=1
for a variety of applications, including agriculture [133], natural
For a determined object class i, the average accuracy APi is ob- resources management [134] and urban planning [135]. To bene-
tained by plotting the model’s Accuracy-Recall curve, it corre- fit from the added value offered by these images, a number of data
sponds to the area under this curve. centers and datasets are publicly available:
• EOSDIS datasets : These are public datasets from twelve Dis-
Note: It is very common to use the mAP50 metric, measuring tributed Active Archive Centers (DAACs), and belonging to the
the mAP value for an IoU threshold of 50%, since it offers a com- Earth Observing System Data and Information System (EOS-
promise between high object location accuracy and error toler- DIS) program of the National Aeronautics and Space Adminis-
ances of this location. tration (NASA), whose mission is to archive, manage and share
various geospatial data. For instance, these data centers include:
3.4.12 EAO (Expected Average Overlap)
– Land Processes Distributed Active Archive Center (LP-
It is a metric assessing the performance of object tracking models DAAC) [136],
on videos. It is calculated by averaging the IOUs obtained on the – Global Hydrology Resource Center DAAC (GHRCDAAC)
image sequence in question [131]. [137],
– National Snow and Ice Data Center Distributed Active
1 N
EAO = ∑ IoU(i), (15) Archive Center (NSIDCDAAC) [138],
N i=1
– Goddard Earth Sciences Data and Information Services Cen-
where: ter (GES DISC) [139],
– Physical Oceanography DAAC (PODAAC) [140].
• N is the number of image sequences making up the video,
These data centers have their own Application Programming In-
• IoU(i) is the IOU of the ith image. terface (API) to ensure seamless data consultation and integra-
tion into customized platforms. Most of these data are remote
3.5 Geospatial datasets sensing images, but other types of data are also available, such
as ground station measurements. Table 2 shows the different
The development of accurate and reliable models depends essen- types of data downloadable free of charge from these data cen-
tially on the quality of training datasets. In addition, the cred- ters, together with their nature and sources.
ibility of the measures used for evaluation also depends on the
• Copernicus Data Space Ecosystem: is a dataset of updated
reliability of the data used. Hence, the need to employ data that is
satellite data taken at different dates, namely data from Sentinel
reliable, normalized, uniform in terms of format and structure and
satellites (from 1 to 5P) and data from Copernicus Contribut-
distributed evenly to ensure that each category of data involved
ing Missions (CCM), implemented in the context of Coperni-
is representative. Furthermore, a dataset must be consistent, i.e.
cus program of the European Space Agency (ESA) [141]. As
rich and diversified, to train the desired model properly with re-
well as EOSDIS datasets, Copernicus Data Space Ecosystem
gard to all possible scenarios. Also, it must be representative of
has its own API, enabling users to freely download and use data
all possible real-life scenarios, so that the trained model is able to
of various sizes and formats [142].
10
 Table 2: EOSDIS main data.

Data source Datacenter Data type Applications

Environmental applications, including monitoring of
Moderate Resolution Imaging Spectroradiometer (MODIS) LPDAAC Satellite images.
vegetation cover, aerosols, land surface.
Ozone Monitoring Instrument (OMI) GES DISC Satellite images. Environmental monitoring and air quality.
A wide range of applications including agriculture and
Landsat satellites (1 to 9) LPDAAC Satellite images.
urban mapping.
Gravity Recovery and Climate Experiment (GRACE) and GRACE
PO DAAC Satellite images. Calculation of terrestrial water reserves.
Follow-on satellites
Tropical Rainfall Measuring Mission (TRMM) satellite GES DISC Satellite images. Precipitation.
Global Food Security-support Analysis Data (GFSAD) LPDAAC Satellite images. Agriculture and food.
AErosol RObotic NETwork (AERONET) GES DISC Ground station data. Aerosol measurement.
Integrated Multi-satellite Retrievals for Global Precipitation Model Hybrid data from several
GES DISC Precipitation.
(IMERG) satellites.
Climate Hazards Group InfraRed Precipitation with Station data Hybrid data (satellite and
GES DISC Precipitation.
(CHIRPS) ground stations).
Water resources management, environment and hydrol-
Surface Water and Ocean Topography satellite (SWOT) PO DAAC Satellite images.
ogy.
Soil Moisture Active Passive satellite (SMAP) NSIDC DAAC Satellite images. Ground humidity measurement.

• Geo-ImageNet: is an extension of ImageNet dataset [143] As well as sattelites data, several geospatial data sources, such as
containing multispectral satellite images and Digital Elevation aerial imagery, LIDAR, trajectory and address data, provide an
Models (DEM) [144]. Geo-ImageNet is developed to train and ideal entry point for training different forms of models. Table 6
validate AI models for geographic object recognition, compris- summarizes the main examples of these datasets, presented in the
ing more than 876 instances spread over 6 object classes. Table remainder of this subsection.
3 shows the distribution of instances according to these classes.
3.5.2 Aerial and UAV images
Table 3: Geo-ImageNet instances.
One of the main advantages of aerial photography and drone im-
Class Instances No.
agery is the very high resolution it offers [171]. Thanks to the low
Basin 155 altitude of acquisition, the images produced enable more robust
Crête 171 information extraction and more powerful analysis capabilities.
Valley 181 Several datasets offer users the possibility of exploiting the per-
Bay 93 formance of these images.
Island 106
Lac 170 • Aerial Image Dataset (AID): as shown in Figure 14, it is an an-
notated dataset of aerial images from different areas around the
world. It is created by Xia et al. [149] to help AI researchers
• xView datasets: these datasets, launched as part of 3 different benchmark the different models to be exploited. AID dataset
challenges, use annotated optical and radar satellite images for contains 10,000 images of 0.5m to 2m resolution belonging to
classification and object detection tasks [145, 146, 147]. Table 30 land use classes, and publicly available in [172].
4 provides detailed informations on each dataset.

In addition to public datasets, several earth observation sattelites

offering high-resolution images have limited access, given their
high accuracy and the sensitivity of some of the areas captured,
since these sattelites have a heliosynchronous orbit, as shown in
Figure 13, which enables them to acquire any geographical area
around the world. Table 5 shows main examples of these sat-
telites.

Figure 14: Examples of AID classes ([149]).

• National Agriculture Imagery Program (NAIP) public

dataset: is a public program launched in 2003 by the United
States Department of Agriculture Farm Service Agency (FSA)
[173]. The purpose of this dataset is to provide free, very high
resolution aerial imagery of agricultural areas in the U.S, pro-
viding a suitable entry point for a variety of applications, such
as land classification and yields estimation [150]. Available
in GeoTIFF format and supported by various platforms, such
as United States Geological Survey (USGS) Earthdata [174],
Google Earth Engine (GEE) [175] or its official portal [150],
NAIP dataset images have a spatial resolution of around 1m
and offer RGB and Near InfraRed (NIR) band images.
Figure 13: Heliosynchronious orbit schema ([148]).

11
 Table 4: xView datasets.

Images Annotations Classes

References Dataset Task Images source
number number number
[145] xView-1 Detection of several object classes. 1127 One Million 60 WorldView-3 30 cm images.
[146] xBD Assessing the damage caused by natural disasters. 22068 850736 6 Maxar open images.
[147] xView-3 SAR Detecting dark fishing. 1000 243018 2 Sentinel-1 images.

Table 5: HR and VHR EO satellites samples.

Satellite Resolution Acquisition bands Organization Manufacturer Launch date

December 17, 2011 (1A) / De-
Pléiades 1A/B 50 cm Panchromatic (0.48 – 0.83 µm) and Multispectral Airbus Defence and Space (ADS)
cember 2, 2012 (1B)
GeoEye-1 41 cm Panchromatic (0.45 – 0.90 µm) and Multispectral General Dynamics September 6, 2008
SkySat satellites (1 November 21, 2013 (SkySat-1),
50 cm Panchromatic (0.45 – 0.90 µm) and Multispectral Skybox Imaging / Planet Labs
to 16) as an example
WorldView-1 50 cm Panchromatic (0.45 – 0.90 µm) Ball Aerospace September 18, 2007
WorldView-2 46 cm Panchromatic (0.45 – 0.80 µm) and Multispectral Ball Aerospace October 8, 2009
WorldView-3 31 cm Panchromatic (0.45 – 0.80 µm) and Multispectral Ball Aerospace August 13, 2014
WorldView-4 31 cm Panchromatic (0.45 – 0.80 µm) and Multispectral Lockheed Martin November 11, 2016
RADARSAT-1 8m C-band SAR MacDonald, Dettwiler and Associates (MDA) November 4, 1995
RADARSAT-2 1m C-band SAR MacDonald, Dettwiler and Associates (MDA) December 14, 2007
ALOS PALSAR-1 10 m L-band SAR Japan Aerospace eXploration Agency (JAXA) January 24, 2006
ALOS PALSAR-2 1m L-band SAR Japan Aerospace eXploration Agency (JAXA) May 24, 2014
Airbus Defence and Space and German
TerraSAR-X 1m X-band SAR June 15, 2007
Aerospace Center (DLR)
Airbus Defence and Space and German
TanDEM-X 1m X-band SAR June 21, 2010
Aerospace Center (DLR)

• University of California Merced (UCM) land use dataset: is

an annotated database of 0.3 m resolution aerial images from
the United States. This dataset, comprising 21 land use classes
such as buildings, airports and parking lots, and totalling 2100
images [151], is used for training and benchmarking classifica-
tion models, and it is accessed via Kaggle platform [176].
• UAVid: is a UAV benchmarking dataset for semantic segmenta-
tion tasks. This annotated dataset, created by Lyu et al. [152] in
2020, totals over 8 classes with 30 videos and 300 drone images,
captured from urban scenes, and it is accessible to the public via
the official website [177].
• UAVSAR: is a public dataset relating to the Uninhabited
Aerial Vehicle Synthetic Aperture Radar program launched by
NASA’s Jet Propulsion Laboratory (JPL) [178], enabling users
to use high-resolution images, from 1 to 3m, through SAR sen-
sors mounted mainly on UAVs, but also on manned aircraft. The
dataset is utilized in a variety of applications, such as landslide
studies, flood monitoring and natural risk assessment [153].
Figure 15: Explanations of the difference between DSM and DTM by horizontal view (a)
[181] and by raster view (b) [182].
3.5.3 Data elevation models
According to Guth et al. [179], Data Elevation Models (DEM) • Copernicus DEM: contains digital surface models (DSMs)
refers to a georeferenced representation of heights at the earth’s from TanDEM-X satellite between 2011 and 2015 [154]. These
surface. As shown in Figure 15, this raster data is representative of DSMs offer a resolution of 90 meters for GLO-30 products and
the bare ground, known as Data Terrain Models (DTM), or of the 30 meters for GLO-90 products, as well as a resolution of 10
surface, otherwise known as Data Surface Models (DSM). DEMs meters only on the European continent for EEA-10 products,
are extracted from optical or radar satellite images, LIDAR point with a good altimetric accuracy Linear error 90% (LE90) of
clouds, or from aerial images, providing a precise topographic 4m. Note that LE90 is the distance within which 90% of linear
representation of the terrain, enabling the construction of 3D vi- altimeter positioning errors lie in relation to their actual position
sualizations, contour lines or slope maps, while facilitating the on the ground [183].
extraction of geographic information [180].

12
 Table 6: Geospatial Datasets with Characteristics and Applications.

Dataset Category Dataset Name Principal Characteristics Applications References

Aerial Image Dataset (AID) Ten thousand images. Object detection and classification tasks. [149]
National Agriculture Imagery Program
Resolution of 1m. Agricultural lands classification. [150]
Aerial and UAV images (NAIP)
University of California Merced (UCM) land
2100 images. Object detection and segmentation tasks. [151]
use dataset.
UAVid 30 videos and 300 images. Segmentation of urban areas. [152]
UAVSAR Resolution of 1 and 3m. Monitoring coastal deformations. [153]
Copernicus DEM Resolution of 90m / 30m. Terrain modelling. [154]
Data Elevation Models ASTER Global DEM Resolution of 30m. Orthorectification. [155]
Shuttle Radar Topography Mission (SRTM) Resolution of 30m. Elevation modelling. [156]
NASA’s Global Ecosystem Dynamics Investi-
Altimetric resolution of 1m. Forest analysis. [157]
gation (GEDI)
USGS 3DEP Resolution of 10 / 30m. Precise 3D modelling. [158]
LIDAR Datasets
UK Environment Agency LIDAR data Resolution of 1m / 50 cm. Precise 3D modelling. [159]
Toronto-3D LIDAR dataset Centimetric resolution. Urban modelling. [160]
NSW Marine LiDAR Topo-Bathy 2018 Resolution of 5m. Hydrographic mapping. [161]
OpenStreetMap Global coverage. Navigation and urban exploration. [162]
Vector Datasets
Google Maps vector data Global coverage. Navigation. [163]
Geolife Trajectories dataset 17621 trajectories. Geolocation and tracking. [164]
Trajectory Datasets ExactEarth Satellite AIS tracking system Global coverage. Maritime traffic. [165]
Multimodal dataset from 3
EuRoC MAV dataset Autonomous vehicle navigation. [166]
sources.
GeoNames 11 million addresses. Geocoding and addressing systems. [167]
Australia Post’s GNAF 13 million addresses. Nationwide geocoding and addressing systems. [168]
Address Datasets
National address database (France) 25 million addresses. Precise nationwide addressing. [169]
ISPARK dataset Istanbul city coverage. Parking lot management. [170]

and topographic correction. SRTM data are publicly available

for 80% of the earth’s surface on various platforms such as Open
Topography [156].

3.5.4 LIDAR datasets

Thanks to their precision, Light Detection And Ranging (LIDAR)
data provides a detailed 3D representation of objects and surfaces
in all directions, through resolved point clouds of up to thousands
of points per square meter. The main LIDAR acquisition cate-
gories are:

• Space-based LIDAR, i.e. on board of sattelites,

• Airborne LIDAR, via aircraft or drones,
• Land-based LIDAR, including mobile or stationary LIDAR,

• Bathymetric LIDAR, operating wavelengths in the green spec-

trum to penetrate water.
Multiple LIDAR datasets with different resolutions are made
available to users to train and to validate models of all kinds:
Figure 16: Extract from a 3 categories copernicus DEM ([154]).
• NASA’s Global Ecosystem Dynamics Investigation (GEDI)
LIDAR : it is an open dataset to assess the world’s forestry
• ASTER Global Data Elevation Model (DEM): is a public structure, using a LIDAR instrument on board of the Interna-
global DTM produced from optical stereo pair images from tional Space Station (ISS) [157]. Each laser pulse has a circu-
NASA’s Advanced Spaceborne Thermal Emission and Reflec- lar footprint of 25m, and the footprints have a spatial sampling
tion Radiometer (ASTER) [184]. This model has a resolution frequency of 60m and an altimetric resolution of 1m. Several
of 30 meters and it is accessible via various platforms such as levels of processing are provided from the raw LIDAR data, for
United States Geological Survey (USGS) earthexplorer [155] instance, Figure 17 illustrates an extraction of GEDI data, corre-
and NASA earthdata portal [185]. sponding to Arial Ground Biomass Density (AGBD), referring
• Shuttle Radar Topography Mission (SRTM): this SAR Band- to the quantity of plant biomass per unit area [186]. This dataset
X mission, combining the efforts of NASA and the German is publicly accessible and it is available in LPDAAC datacenter
space agency, aims to produce a global DTM. As of 2014, the [136].
resulting model is publicly available in 30-meter resolution, and
it is used in numerous applications such as orthorectification
13
 • Microsoft Global Building Footprint : is a collection of vec-
torial building footprints extracted from satellite images using
semantic segmentation, achieving an average accuracy of 90%
[189]. It is a vast collection of over 1.3B footprints across 5 con-
tinents, using images derived from Maxar [190], Airbus [191]
and IGN France [192]. Figure 19 illustrates a concrete example
of this data.

Figure 17: Footprints of AGBD values (in Mg ha-1), acquired from April to July 2019 ([187]).

• USGS 3DEP LIDAR data : is a USGS open dataset shared in

the context of 3D Elevation Program (3DEP), the aim of this
program is to provide users with 3D airborne LIDAR data for
the whole of the United States [158]. This dataset is available
at a resolution of 30 m for rural areas and 10 m for cities [188].
Figure 19: Microsoft building footprints samples ([189]).
• The UK Environment Agency LIDAR Data : this public
dataset contains airborne-LIDAR data at the scale of England,
with a spatial resolution of 1m, enhanced to 0.5m for the dens- • OpenStreetMap : this is vector data from the OpenStreetmap
est areas, as well as derived products, i.e. DSMs and DTMs collaborative mapping project, founded in 2004. The goal of
[159]. this project is to provide a public representation of the world
according to the land use in three standard vector forms, i.e.
• Toronto-3D LIDAR dataset : is an annotated dataset compris- points such as restaurants and hotels, lines including roads and
ing mobile terrestrial LIDAR data to train semantic segmenta- waterways, and polygons for instance parks and lakes [193].
tion and classification models. This dataset features centimet- This data is downloadable from the official Openstreetmap web-
ric horizontal and vertical resolution, while capturing the urban site [162], via multiple platforms such as Geofabrick [194], and
roads of the city of Toronto in Canada [160], an example of BBBike [195].
which is shown in Figure 18.

Figure 18: Example of Toronto-3D LIDAR dataset ([160]).

• NSW Marine LiDAR Topo-Bathy 2018 : This dataset com-

bines airborne topographic radar and bathymetric LIDAR, cov-
ering the coastal areas of New South Wales (NSW), Australia.
With a horizontal resolution of 5m and vertical resolution in the Figure 20: OSM roads network, example from Morocco, updated 08/23/2024.
decimeter range, this dataset enables precise coastal mapping,
and serves as an ideal entry point for training mapping models
[161]. • Google maps vector data : as part of its mapping services, i.e.
Google maps [163] and Google earth [196], Google offers vec-
tor tiles representing various geographical features, including
3.5.5 Vector data roads, buildings, etc. Although this data is protected by user
Vector data, as digital representations of geographic reality in the rights, due to the cost of its collection and production, it can
three standard geometric forms of point, line and polygon, pro- be used via Google Maps APIs, such as Places API [197] and
vide a rich source of information while facilitating the geospatial Directions API [198].
analysis process. The reasonable size of these data in memory and
the possibility of extracting them from several sources, including 3.5.6 Trajectory datasets
images and maps, encourage users to employ them in various ap-
plications, among which vector geospatial datasets: Trajectory data, such as Global Navigation Satellite System
(GNSS) and Inertial Navigation System (INS) data, enable the
14
precise location of objects and people. This is particularly useful
for analyzing the behavior of human groups and the flow of re-
sources and energy over time and space. In this subsection, the
main examples of trajectory datasets are presented.
• Geolife Trajectories dataset: is a set of GPS trajectories from
more than 178 users in over 30 Chinese cities. Collected by Mi-
crosoft Research [199], this dataset contains a total of 17621 tra-
jectories, covering 1251654 Km over a period of 48203 hours.
Figure 21 shows the density of trajectories recorded in the city
of Beijing, also, it is noted that this dataset is freely available on Figure 22: Geonames density map ([201]).
Kaggle platform [164].
• Australia Post’s Geocoded National Address File (GNAF)
: is a public database containing over 13 million addresses in
Australia. This dataset is characterized by two elements: geo-
graphic coordinates and location labeling according to the Aus-
tralian administrative system. GNAF is freely accessible via
[168].
• National address database : The French government have
launched a collaborative project between the Interministerial
Digital Department [202], the National Agency for Territorial
Cohesion [203] and the National Geographic Institute [204], in
order to provide the public with a geolocated, daily-updated na-
tional address database. This open database, as illustrated in
Figure 23, is available in several formats, containing over 25
million addresses, and it is downloadable from [169].

Figure 21: Geographical extent of GeoLife dataset for the city of Beijing.

• European Robotics Challenges - Micro Aerial Vehicles (Eu-

RoC MAV): is a public dataset containing mainly inertial nav-
igation system (INS) data, acquired from inertial measurement
unit (IMU) sensors such as accelerometers and gyroscopes, as
well as GPS trajectories and stereoscopic images, from micro-
aerial vehicles (MAVs). The aim of this dataset is to analyze
the behavior of these vehicles, in order to further research in the
context of associated navigation, localization and operational
models [166].
• ExactEarth Satellite AIS tracking system: is a private dataset
from the satellite-based Automatic Identification System (AIS)
[200]. Using radio frequency (RF) waves, ships can trans-
mit their GPS position, speed and heading, providing real-time Figure 23: French adress database sample from Paris city, OpenStreetMap basemap.
global data on the trajectory of maritime traffic through the AIS
system [165].
• ISPARK dataset : is a public dataset created by the munici-
pality of Istanbul, Turkey, containing all the city’s parking lot
3.5.7 Adress datasets addresses, their capacities and the associated tariffs [170].
These datasets correspond to a set of information about specific
An essential point to raise is the relevance of public-access
geographical locations, allowing to pinpoint their precise loca-
datasets, given that sharing data helps to improve its scope and
tion and describe them in detail. This is particularly important
reliability, especially in the case of collaborative datasets such as
when an analysis of locations is suggested, such as collecting cus-
Openstreetmap [162] and Geonames [167]. Moreover, the trans-
tomer reviews, responding to emergency alerts, managing water
parency of these datasets enables their reliability to be measured
and electricity services, etc. Examples of such datasets include
autonomously, via a variety of third parties and methods. In re-
the following:
turn, the use of these data in a wide range of research projects
• GeoNames : is a geographic database containing over 11 mil- validates the results of these studies, guaranteeing the scientific
lion addresses worldwide, including geographic coordinates, community a reference base from which to compare the studies
zip codes, population and more. These data, whose global den- carried out. In the next subsection, the hardware aspect is ex-
sity is illustrated in Figure 22, are available under [167]. plored, including its different categories and impact on the perfor-
mance of various types of models, as well as the latest hardware
infrastructures designed for optimal AI efficiency.

15
3.6 Hardware data, while enabling the data used to be stored, accessed and ma-
nipulated quickly enough. There is also a strong need to optimize
Pioneering progress in the development of hardware architectures
hardware architectures, as this naturally implies improving GeoAI
and components has boosted the performance and efficiency of
model training and inference processes, both in terms of time and
AI, both in terms of training models and predicting results for
energy consumption.
different cases. An understanding of available hardware opportu-
nities is therefore necessary to choose the right hardware config-
uration for a given AI task. In this context, a range of hardware 4 GeoAI applications
components involved in these processes are distinguished, includ-
ing hardware accelerators, defined as hardware performing a spe- Having identified research questions and corresponding research
cial function faster than CPUs. To illustrate, numerous hardware directions, this section paves the way for a deeper understand-
accelerators are cited: ing of GeoAI applications for precision agriculture, environment
and natural disaster monitoring, water resource management, ur-
• Graphics Processing Unit (GPU) [205],
ban planning and healthcare. Figure 24 shows the details of these
• Field-Programmable Gate Array (FPGA) [206], research axes and sub-axes. Indeed, the theme of artificial intel-
ligence applied to geospatiality is addressed not only as an asset
• Digital Signal Processor (DSP) [207], for processing and manipulating complex data, but also as a pow-
• Tensor Processing Unit (TPU) [208], erful and advanced decision-making tool that is methodological,
effective and efficient.
• Application-Specific Integrated Circuit (ASIC) [209].
In addition to these advanced accelerators, there are storage de-
vices such as Solid-State Drives (SSDs) [210], revolutionizing the
optimization and performance of data entry and access, together
with networking hardware, useful in cases where, for example,
numerous users request predictions of a given model.
Several researchers have carried out comparative studies of hard-
ware implementation of AI algorithms. For example, Abu Talib
et al. [211] provided a systematic review comparing the perfor-
mance of a multitude hardware accelerators when running DNNs,
especially FPGA, GPU and ASIC, and concluded that FGPAs pro-
vide a high reconfiguration capacity, despite their limitation in
terms of computing power. GPUs, on the other hand, demon-
strated a high execution capacity thanks to their parallel architec-
ture, optimising high computation operations. However, ASICs
show high efficiency in specific tasks, for instance video process-
ing on mobile devices, object recognition using embedded sys-
Figure 24: Research axes for GeoAI.
tems, while demonstrating significant energy optimization.
In a different context, Dally et al. [212] have taked the exam-
ple of CNNs to examine the requirements of the basic operations 4.1 Precision agriculture
involved, essentially convolution and matrix multiplication. For
training operations, multiplication operations require 16-bit float- According to Pierce and Nowak [216], precision agriculture refers
ing point precision and results are summed using 32 bits, making to methods that use new technology to manage different aspects
computational precision a priority in the design of electronic cir- of agricultural production in order to improve productivity, boost
cuits. For prediction operations, a much lower precision (8 bits) is crop yields and control the associated environmental quality. Con-
more than sufficient to perform calculations, making energy opti- sequently, the close collaboration between precision agriculture
mization and time reduction a priority in this case. This difference and GeoAI has led to a better understanding of the needs and chal-
in design priorities for each type of operation makes it necessary lenges of the agricultural sector, enabling to offer cutting-edge
to optimize them in order to adapt to all requirements, i.e. energy support, whether through precision mapping at the finest scales,
savings, speed and calculations efficiency. Additionally, Zhao et yield prediction or precision irrigation. Table 7 shows the main
al. [213] present Techtonic-shift, a novel infrastructure to im- methods and models explored in this field.
prove I/O storage request procedures for ML training tasks, while
reducing energy demand. The proposed infrastructure combines 4.1.1 Crops mapping
Tectonic large-scale distributed file systems [214] with integrated
It is clear that the integration of artificial intelligence methods
Flash storage [215], while introducing memory cache manage-
into the process of mapping agricultural yields, thereby produc-
ment protocols in order to optimize invloved demands.
ing added value geographical data, enables significant progress to
Given the size and diversity of geospatial data, GeoAI requires
be made in detecting the key factors influencing yields, including
high-performance hardware to ensure that the data in question
land preparation and fertilisation, crop protection against perti-
can be trained effectively. The use of accelerators such as GPUs
cides, the automation of tedious tasks, the identification of large-
, FPGAs and TPUs with considerable RAMs is recommended
scale crop types and the effective planning of targeted activities.
to perform the concerned tasks. For example, GPUs are per-
Relevant studies successed to achieve a high level of control while
fectly suited for computer vision tasks such as remotely sensed
using different categories of AI models and geospatial data. Han
images segmentation, while FPGAs are more suited for GeoAI
et al. [223] presented an innovative spatio-temporal multi-level
tasks that require real-time processing, such as vehicle tracking
attention (STMA) model for crop mapping using Sentinel 1 SAR
and AIS data census. In addition, storage devices, whether lo-
images, including three datasets from Germany [224], France
cal or in the cloud, must take into account the huge volume of
16
 Table 7: Proposed AI methods for precision agriculture.

Applications Methods and algorithms References

Mask Region-based Convolutional Neural Network combined with Random Forest, Support Vector Ma-
[217]
Crops mapping and classification chine and Multiple Linear Regression.
Linear mixed-effect. [218]
Random forest. [219]
Yields prediction
Random Forest and boosting methods. [220]
Mask Region-based Convolutional Neural Network and ResNext-101. [221]
Precision irrigation
U-Net architecture. [222]

[225], and South Africa [226], covering various crop types. The mapping, particularly in small-scale study areas. Ribero et al.
proposed model includes ResNet network [86] for feature extrac- [231] evaluated three different traditional convolution-based net-
tion, a cross-attention mechanism to regularize data dimensions, a works, specifically U-Net [85], LinkNet [232] and Pyramid Scene
spatio-temporal self-attention module to extract relationships be- Parsing Network (PSPNet) [233] for the automatic segmentation
tween spatial and temporal dimensions and a decoder based on of crop rows in sugarcane fields. To train and evaluate the models,
U-Net [85] architecture to reconstruct the feature map. Overall the authors used senseFly Sensor Optimized For Drone Applica-
Accuracy (OA) and F1-Score metrics are used during the evalua- tion (SODA) images, an example of which is shown in Figure 27.
tion phase, obtaining average values across all classes of 96% and Results of validation on four samples showed marked stability,
74% respectively. with U-Net network performing relatively better than the other
two models in terms of Dice coefficient, being a statistical sim-
ilarity metric [234], used in this context to assess segmentation
capacity. Table 8 shows the Dice coefficient values on the said
samples.

Table 8: Comparison between U-Net, LinkNet and PSPNet on SODA dataset ([231]).

Sample U-Net LinkNet PSPNet

A 90.8% 89.1% 87.1%
B 92.3% 90.3% 88.6%
C 90.4% 88.4% 86.6%
D 86.7% 83.6% 84.1%

Figure 25: Examples of SAR Sentinel1 images used in [223].

A classificaton model based on satellite imagery is proposed by

Gallo et al. [17] for the automatic mapping of seasonal crops
through Sentinel-2 images along with old crop maps. The con-
cerned model is composed of two parts, a 3D Feature Pyramid
Network (FPN3D), which is a 3D extension of Feature Pyramid
Figure 27: Samples of data used for crops segmentation ([231]).
Network (FPN) [227], and an aggregator CNN allowing annual
segmentation. As shown in Figure 26, training dataset is con-
structed using phenological analysis to enhance historical crop Research works cited in this subsection has demonstrated the ca-
maps, as well as recent Sentinel-2 satellite imagery. FPN3D pabilities of GeoAI for predictive mapping of agricultural crops,
model is then generalized through a CNN aggregator to produce including the use of satellite imagery at different resolutions, his-
an accurate seasonal map of agricultural crops. A test phase is torical crop maps for fairly accurate classification, and ground
carried out while obtaining an overall accuracy (OA) of 73% over truth data for validating the performance of the AI models in ques-
four successive seasons. Tiwari et al [228] combined Sentinel- tion. This mapping technique facilitates the evaluation of the im-
1 SAR and Sentinel-2 optical imagery to identify the transplation pact of drought and climate change on agricultural production,
and peak seasons of rice crops, using two approaches, random for- along with the identification of resulting yields, this is of direct
est classification and multi-Otsu approach [229]. Results showed relevance to the next subsection.
a classification accuracy of up to 94%. Following the same logic,
Mohammadi et al. [230] used time series of Landsat 7 and 8 data, 4.1.2 Yields prediciton
as well as old geographic yield layers to improve the efficiency
and accuracy of cartographies, especially when the model is con- Several studies exploited advanced AI techniques to develop ac-
fronted to unseen data. The approach is to supervise the proposed curate models for yields prediction. Carneiro et al. [219] used
3D Fully Convolutional Network (FCN) layers to understand the a random forest classifier to identify the most significant factors
spatio-temporal relationships between data series. An evaluation in predicting cotton yield. Additionally, a random forest regres-
carried out on unseen samples demonstrates a high classification sor is used to generalize this prediction on the basis of the clas-
performance, with an average IOU of 81% and an F1-Score of sifier results. The authors have exploited data of various types
90%. and have shown that indicators derived from Sentinel-2 satellite
Apart from satellite imagery, several studies explored the impor- images, real-time kinematic (RTK) GPS and LIDAR data are the
tance of high-resolution UAV imagery, for more accurate crop most decisive. Subsequently, results showed an R2 of 62%, an

17
 Figure 26: Workflow for the automatic mapping of seasonal crops ([17]).

MAE of 0.16 and a MAPE of 5% for the five most relevant factors. This subsection gives, through the cited papers, a global overview
Ramzan et al. [220] have combined the exploitation of Landsat-8 of the latest GeoAI techniques for the quantification and estima-
satellite data converted to Normalized Difference Vegetation In- tion of agricultural yields. Different Machine Learning and Deep
dex (NDVI) [235], a vegetation index using red and near-infrared learning models contribute in the said task using various geospa-
(NIR) bands to measure vegetation density on a given surface, tial data, mainly satellite and UAV images, proving to be very
along with agro-meteorological data, to train a multi-modal AI interesting for accurate and reliable prediction of large-scale agri-
model. In this way, the authors use MLP, support vector regres- cultural productivity.
sion [236] and Gaussian random projection, referring to a dimen-
suality reduction technique based on matrices whose elements fol- 4.1.3 Precision irrigation
low a Gaussian distribution [237]. Ensemble learning methods are
also employed, specifically random forest as a bagging method, Abioye et al. [241] defined precision irrigation as the optimal use
extreme gradient boosting and XGBoost as boosting methods. Fi- of water resources through the introduction of modern technolo-
nally, a DNN with 3 hidden layers is used to estimate tea yield gies in the agricultural procedures considered. Indeed, the use
at farm scale. The resulting model proved its performance by of prediction performance offered by GeoAI methods in new ir-
achieving an R2 of 99%. rigation technologies is becoming more applicable and efficient.
Given the high data accuracy provided by UAV imagery, several In this context, de Albuquerque et al. [221] used a Mask-RCNN
works address yield estimation using UAV data. Peng et al. [217] [238] with ResNeXt-101 backbone [242] for automatic detection
used an original approach to conveniently and economically esti- of Center Pivot Irrigation Systems (CPIS), donating automatic
mate wheat crop yields by using UAV images, example of which watering systems for large agricultural areas by rotating large
are shown in Figure 28. The first goal of this study is to extract the pipelines around a central point. The global model takes as input
phenotypic characteristics of wheat spikes, using a Region-based time series of Sentinel-2 satellite images with data augmentation
Convolutional Neural Network (R-CNN) [238]. Then, the yield as input, besides, it is conceptualized to take into account differ-
is estimated using classical ML algorithms, specifically Random ent environmental situations, including clouds and seasonal vari-
Forest Regression (RFR), Multiple Linear Regression (MLR) and ations. While averaging the simulations results over six tests, the
Support Vector Regression (SVR). A test phase is executed, re- model reached a mAP50 value of 88%. Raei et al. [222] presented
sulting in an F1-Score of 83% for the Mask-RCNN. In addition, an U-Net segmentation model [85] with ResNet30 backbone [86],
R2 coefficient is used to examine the performance of RFR, MLR combined with a transfer learning approach for irrigation systems
and SVR models, obtaining values of 83%, 85% and 86% re- segmentation on a regional scale. 8 600 aerial images from NAIP
spectively. In the same context, Wang et al. [218] used a Lin- dataset [150] are used to train, evaluate and test this model. There-
ear Mixed-Effect (LME) model, referring to a regression model fore, final results show a high accuracy of up to 94% while prov-
considering the relationships between the independent variables ing the use of transfer learning and the unbalanced dataset of sev-
and a dependent variable, modelling the intrinsic variability of eral irrigation systems categories.
the data by means of an intrinsic term in the regression equation In addition to the automatic detection of irrigation systems, sev-
[239]. The ultimate aim of this research is to accurately estimate eral studies tackled the subject of optimizing water resources ded-
the leaf area index (LAI) [240] of rice crop. Diverse ML mod- icated to irrigation. Jalajamony et al. [243] proposed an intelli-
els are used to compare the performances obtained, in particular gent system enabling selective irrigation of the driest areas while
simple regression (SR), artificial neural network (ANN) and RF optimizing consumption. This system uses several data sources,
classifier and regressor, while using RGB UAV images and rice including RGB images integrated into the Raspberry Pi 4, Ther-
variety informations. The results showed that LME model is the mal InfraRed (TIR) images from the Lepton 3.5 camera, under
most accurate, achieving an R2 of 76% to 81% and an RMSE of different levels of hydration, so as to visualize the behavior of the
1.04 to 1.16. thermal camera in relation to this application. In addition, Global
Positioning System (GPS) data from a customized quadricopter
is employed to geolocate the areas to be used. Moreover, an AI
component is implemented to exploit the embedded results by ad-
justing the angle of rotation of the flow control valve, this com-
ponent uses Random Forest (RF) as a base model resulting in an
MSE of 0.06, compared to values of 0.14 and 86.5 for KNN and
SVM respectively.

Figure 28: Extractions of an UAV image used for estimating wheat crop yields ([217]).

18
Figure 29: Examples of thermal images under different levels of hydration (a) and at different
altitudes (b) ([243]).

Despite the complexity of irrigation procedures, GeoAI has

proven to be a high-performance tool for precision irrigation. This
includes the detection of irrigation systems from remote sensing
images, in addition to the use of several models as a key compo- Figure 30: Examples of UCM dataset ([151]).
nent for the implementation of concerned systems and procedures.
These same models allow for equally effective contributions to
As well as classification, a number of studies have focused on cre-
the field of urban planning, logistics and transport, enabling sus-
ative methods for building extraction from geographic data. For
tainable and optimal management of the densest urban areas, and
instance, a U-Net architecture [85] is used by Zhang et al. [246]
constituting the problem to be explored meticulously in the next
for 3D buildings extraction from LIDAR data and photogrammet-
subsection.
ric point clouds. This approach extracts the shapes of real colored
buildings with high efficiency, obtaining an overall accuracy of
4.2 Urban planning, logistics and transport 87%.
Through the census implemented in this subsection, it is clear that
By combining AI models with geographic information acquisi-
the extraction of highly filtered data from multiple datasets rep-
tion techniques, urban planners can better understand and pre-
resents a topical issue. In addition, it is remarked that multiple
dict possible trends in urban dynamics. This logic is perfectly
categories of geospatial data are used, including aerial, satellite
synchronized with the integration of new technologies into urban
imagery and LIDAR point clouds, while noting the particular con-
management tools, while allowing researchers to implement in-
tribution of resolved data in classification and segmentation tasks,
novative models that enables the inhabitants of major cities to live
given the inherent need for much finer data extraction.
”better”. In the remainder of this subsection, relevant works com-
bining GeoAI and urban dynamics, smart-cities modelling, logis-
tics and transport along with the management of green urban areas 4.2.2 Urban dynamics
are listed. Table 9 summarizes these research studies. Thanks to their ability to process complex data ”intelligently”,
new ML and DL techniques are able to completely change the
4.2.1 Urban data extraction classical understanding of urban dynamics, enabling the anticipa-
tion of the main needs on this field. Tehrani et al. [247] proposed
Urban classification, i.e. the categorisation of existing input data,
a novel ANN to predict the energy of solat electromagnetic radia-
such as remotely sensed images or scanned maps, into several
tions in urban areas, using a fruitful combination of different types
classes, enables urban data to be analysed and grouped with a
of data, precisely three-dimensional coordinates of constructions,
high degree of precision, taking into account their complexity
average building heights, inhabited and unoccupied areas coupled
and diversification. Alem and Koumar [244] used UCM dataset
with azimuth angle. This study constitutes a creative work to in-
[151], to compare three classification models, comprising a Con-
tegrate solar energy into future and sustainable urban planning.
volutional Network for Feature Extraction (CNN-FE) along with
Simulations results showed that this model achieves an MSE of
a Transfer Learning (TL) and a fine tuning models, adjusted
0.01 and an R2 of 85%. Hajjar et al. [248] proposed an LSTM
from a pre-trained EfficientNet model [262]. Performance met-
to process Land-Use Change (LUC) data in time and space. Us-
rics showed that the fine tuning model performed better in terms
ing 3 raster maps with dates of 2000, 2010 and 2020, the tests
of classification, achieving an accuracy of 88%, and an F1-Score
have showed significant effectiveness for the classification task in
of 89% for 21 classes. In the same context, Chen et al. [245] pro-
question.
posed a DL model to detect urban villages in the main cities of
In a relatively different context, prediction of urban development
Jing-Jin-Ji region from China. The input data consists of Google
is a key element in the preparation of urban master plans, as they
Earth images merged with Street View data [263]. This model has
indicate recommendations to be followed in the future. Sun et al.
two branches, a neural network based on multi-scale dilated con-
[249] developed a pix2pix GAN [266] for Land-Use and Land-
volutions [264] for the Google images and a Recurrent Attention
Cover Change (LUCC) prediction by integrating vectorial plan-
Network based on multi-view learning [265] for the Street View
ning indicators, as well as land-use probability maps from Landsat
data, while achieving an accuracy of 92%.
5 and 8 satellite images. Model in focus performed well, achiev-
ing a pixel-by-pixel prediction accuracy of 85%.
To conclude, various models are used to explore the different
aspects of urban dynamics, including the distribution of energy
over urban environments, the analysis of spatio-temporal changes,
19
 Table 9: Proposed methods for urban planning.

Applications Methods and algorithms References

Convolutional Neural Network - Feature Extraction. [244]
Urban data extraction Multi-Scale Dilated Convolutional Neural Network and Multi-Scale Dilated Convolutional Neural Network. [245]
U-Net model. [246]
Artificial Neural Network. [247]
Urban dynamics Long Short-Term Memory model. [248]
Pixel to pixel Generative Adversarial Network. [249]
MPSiamese, Few-Shot Siamese Networks and YOLOv5-Pyramid model. [250]
Smart cities Long Short-Term Memory model. [251]
Random Forest classifier, Xgboost and Faster Region-based CNN. [252]
Support Vector Machine, Feedforward Neural Network, Convolutional Neural Network, Recurrent Neural Network, and
[253]
Long Short-Term Memory.
Hybrid Gated Recurrent Unit - Long Short-Term Memory. [254]
Logistics and transport Incremental Output Decomposition Recurrent Neural Network. [255]
Denoising stacked Autoencoder and a Collaborative Learning approach. [256]
SegNet model. [257]
Multi-modal deep learning model. [258]
Semi supervised CNN. [259]
Green infrastructures and Bio-security Faster Region based CNN. [260]
Convolutional Neural Networks. [261]

whether in relation to land cover, linking to the occupation of

the urban surface in focus, or in relation to land use. In short,
GeoAI techniques are a key tool for monitoring and analyzing ur-
ban changes, and for launching a new era of intelligent cities, a
theme analyzing in depth in the following subsection.

4.2.3 Smart cities

The integration of cutting-edge AI models in smart cities is in-
creasingly indispensable to automate urban infrastructure. Indeed,
a multitude of research projects are underway to combine geospa-
tial data with AI techniques In this regard. Hu et al. [250] pre-
sented an intelligent spatio-temporal framework based on GeoAI.
The proposed framework integrates multi-video geo-referenced
approach along with the virtual city based on Augmented Reality
Figure 31: Examples of Street View images used in [252]. The blue boxes represent road
(AR) in urban digital twins. Siamese [267] video tracking models signs and the red ones shop signs.
are utilized, in particular MPSiam [268] and Four-Stream Siamese
Network (FSSiamese) [269], together with YOLOv5-Pyramid ar-
In conclusion, this subsection provides a complete overview of
chitecture [270] for small object detection. The presented models
the contribution of GeoAI to the development of smart cities. The
demonstrated significant performance while obtaining a mean off-
synergy between AI models and geospatial data makes it possible
set error for video georeferencing of 0.14m. In a different context,
to exploit these data in a multitude of task, such as detecting key
Canli and Toclu [251] have developed a mobile application based
objects, exploring the availability and status of the emphasized
on deep and cloud learning to refine parking searches in the city of
services, and improving the attractiveness of commercial offers.
Istanbul. The authors used an LSTM fed by the ISPARK dataset
collected by Istanbul Metropolitan Municipality is used in this
context. Test results showed the high efficiency of LSTM model 4.2.4 Logistics and transport
in predicting car park occupancy rates, compared with RF, SVM In addition to urban classification and the identification of the con-
and Autoregressive Integrated Moving Average (ARIMA) mod- cerned dynamics, GeoAI provides productive solutions for opti-
els. The parking searches model achieves an accuracy of 99% mizing and forecasting traffic, calculating routes and managing
and an RMSE of 1.59. various transportation systems. Several AI models using geo-
Ye et al. [252] introduced the factor of urban commerce in smart graphic records are employed to improve the safety of road net-
cities by analyzing Street View annotated images, illustrated in works, to ensure the sustainability of transportation modes and to
Figure 30, and OpenStreetMap road network. Random Forest and manage logistics operations. In this settings, Panovski and Za-
Xgboost classifiers are applied to the images and a Faster Region- haria [253] have proposed a method for predicting bus arrival
based CNN (Faster-RCNN) model [271] is employed to detect times in short term, i.e. from a few minutes to two hours, and
commercial signs. Results showed satisfactory accuracy in differ- in long term, i.e. throughout the day, using a traffic density ma-
ent scenarios, while obtaining an average precision (AP) of 88% trix (TDM). As part of a comparative study, numerous models are
for the detection task and a mAP of 85% for Random Forest clas- used including Support Vector Regression (SVR), CNN, Feedfor-
sification. ward Fully Connected Neural Network (FNN), RNN and LSTM
to improve the accuracy of predictions, giving a Mean Absolute
Error (MAE) values of 71 s, 48.7 s, 46.9 s and 45.3 s, 44.6 s for
the SVR, CNN, RNN, LSTM and FNN respectively. Likewise,
20
Zafar et al. [254] have developed a hybrid Gated Recurrent Unit 86% and an F1 score of 74%.
LSTM (GRU-LSTM) for predicting traffic in smart cities. The
input data is various OSM data and GIS maps for road segmenta-
tion. The model is 95% accurate. Lu et al. [255] proposed an In-
cremental Output Decomposition Recurrent Neural Network (IO-
DRNN), referring to a specific RNN based on incremental learn-
ing, for traffic prediction in Intelligent Transport Systems (ITS),
attempting to overcome the shortcoming of time delay in predic-
tions. GNSS data is employed as input to the model, with an MAE
and an RMSE of 16% and 17% respectively.
In addition to predicting traffic density, route calculation aims to
reduce the number of trips and the consistency of congestion, con-
stituting a field of attack for GeoAI tasks. Chen et al. [256] pro-
posed a DL model for recommending optimized itineraries for
tourists, while integrating points Of interest (POI) location with
Figure 33: Sample of rasterized LIDAR data used for railway features segmentation ([258]).
user history. The concerned model is composed of a Denoising
Stacked AutoEncoder (SDAE) [272] and a collaborative learn-
ing model [273] to predict user interests and visit durations. As Through the inventory presented above, various researches have
shown in Figure 32, the SDAE architecture consists of an encoder exploited GeoAI models for traffic management, route optimisa-
that learns the robust representations of noisy textual descriptions tion and inventory of existing infrastructure, enabling decision-
of POIs, while reducing the associated dimensuality. The de- makers to improve the accuracy of their estimates and to imple-
coder then reconstructs the original dimension, while correcting ment high-calibre strategic planning. In the next subsection, a
the noise introduced. The model results in an average F1-score rather interesting area is explored, in particular that of green in-
of 48%, obtained over 8 different cities, namely Budapest, Delhi, frastructure and biodiversity and their impact on a healthy urban
Edinburgh, Glasgow, Osaka, Perth, Toronto and Vienna. The per- environment.
formance obtained is considered very favourable given the com-
plexity of the task in question. 4.2.5 Green Infrastructure and Bio security
Optimal management of green spaces in urban areas is crucial
to enhance the urban landscape and to improve significantly the
quality of life of city residents. Many research studies have con-
tributed to this perspective. For instance, Guo et al. [259] pro-
posed a semi-supervised CNN based on Deeplabv3+ architec-
ture [277] for the Urban Tree Cover (UTC) mapping. Images
from Digital Globe’s WorldView-2 and 3, GeoEye-1, Planet Labs’
SkySat and Airbus’ Pleiades satellites are adopted to train the
model. Evaluation phase showed encouraging results, with an
overall accuracy of 95%. Furthermore, an individual tree inven-
tory (ITI) model is developed by Das et al. [260] to quantify
the contribution of tree cooling on urban heat. Aerial high res-
olution and Landsat-8 Land Surface Temperature (LST) images
together with LIDAR data are utilized to describe the Canopy
Figure 32: SDAE architecture for route calculation, proposed in [256]. Height Model (CHM) with a resolution of 1m. The proposed
model architecture is based on a Faster Region based CNN (R-
The inventory of transportation infrastructure seems necessary for CNN) architecture [271] coupled with Inception ResNet V2 [278]
quickly identifying areas in need of maintenance, and to quickly in order to identify seven individual tree canopy species. Addi-
update road conditions. Rúa et al. [274] estimated the road areas tionally, Pearson correlation [279], Linear Regression (LR) and
likely to be affected by rockfalls, while analyzing the slopes of Geographically Weighted Regression (GWR) [280] are deployed
roads. The authors first detect the slope through a cloud of LIDAR to identify the relationship between tree characteristics, species
points, estimate the areas susceptible to falls, and then simulate and local temperature. The resulting identification model per-
the event using RockGIS software. Indeed, the F1-Score of the formed very well, achieving a mean Average Precision (mAP) of
Proposed model is 96%. In addition, Kearney et al. [257] have 85% and an average F1-Score of 80%.
presented a segmentation model based on SegNet [275], an en- In addition to inventorying green spaces in urban environments,
coder decoder architecture based on CNN, using rapidEye satel- Carnegie et al. [261] have developed a tree classification model
lite images with the aim of updating rural road networks for forest for forest pest biosecurity monitoring. The model uses two CNNs
landscapes. The test phase showed high capabilities, with a Re- for tree classification (M1) and instance segmentation (M2), with
call of 89% and accuracy of 87%. In another context, Grandio et multispectral aerial imagery as input. Results of this comparison
al. [258] presented a multimodal DL model for panoptic segmen- favoured Model M2 over M1 in terms of F1-Score metric, attain-
tation of railway features, including signposts, rails, information ing an average value of 79%, in comparison with a value of 58%
boards, etc. A point cloud from two LIDAR sensors is utilized for M1.
to generate railway infrastructures. As illustrated in Figure 33, In summary, various ML and DL models are used to optimize
a U-Net architecture [85] is applied to rasterize the said point the location of green infrastructure and to guarantee bio-diversity
cloud. Subsequently, a Pointnet++ [276] model is exploited for in urban environments, thereby ensuring the ecological balance
a detailed semantic segregation. This workflow have showed re- of the environment in question, improving the quality of citizens
markable performance, achieving a precision of 68%, a recall of life and accurately monitoring urban trends. Cited Research work

21
includes urban vegetation mapping, correlation measurements be- [296] presented an improved leaf area index at 250 m resolution.
tween green infrastructure and environmental parameters, as well For this, the authors used a bidirectional long-term memory (Bi-
as monitoring biodiversity within the most densest cities. The LSTM) [297] as a base model. MODIS surface reflectance data
control of green infrastructures and biodiversity leads to focus on is exploited to form a spatio-temporal series as a training dataset,
another topic, that of environmental monitoring, comprising cli- while producing 79 LAI maps to test the resulting model.
mate change, atmospheric chemistry and environmental impacts. Despite the difficulties associated with managing the problem of
climate change, several AI models using geospatial data are im-
4.3 Environmental monitoring plemented to monitor, analyze and predict the causes and conse-
quences of this phenomenon. Accurate and comprehensive anal-
Complex data from a variety of sources such as satellite imagery, ysis is carried out to manage the resulting effects and to plan ap-
environmental sensors and climate variables, can be used to mon- propriate coping strategies.
itor critical aspects of the environment, such as climate change
measurements, air quality variables and environmental impacts of 4.3.2 Atmospheric chemistry
projects, activities and events. In fact, Figure 34 shows an exam-
ple of environmental data exploited in this context, more precisely Atmospheric chemistry is a sub-discipline of atmospheric sci-
the Aerosol Optical Depth (AOD), referring to the portion of light ence quantifying the composition of chemical elements in the at-
attenuated by aerosol particles in the atmosphere [281], measured mosphere. Besides, artificial intelligence, in particular GeoAI,
by OMI instrument and downloadable via EOSDIS datasets [142]. guarantees accurate prediction of atmospheric chemical compo-
Conversely, Table 10 summarizes AI methods explored in this nents, through training based on historical data. Opio et al. [285]
subsection. have developed a Weather Research and Forecasting model for
the atmospherical Chimestry (WRF-Chem) based principally on
OMI images. In addition, the proposed architecture has been
enhanced by a deep convolutional autoencoder for WRF (WRF-
DCA) to model the atmospheric dispersion of sulphur dioxide
(SO2). This model showed a strong performance with an RMSE
of 1.5 × 1016 molecules/cm2 . Kim et al. [286] used an XG-
Boost model to estimate the hourly Nitrogen Dioxide gaz (NO2)
concentrations from Sentinel 5P images, along with meteoro-
logical data and ground sensor measurements. Results showing
the importance of accurate near-surface NO2 mapping, particu-
larly during the COVID-19 containment period. For the same
reason, Dou et al. [287] have combined an RF model with K-
Figure 34: OMI Near UV Aerosol Optical Depth dated 07/24/2024.
Means algorithm to predict NO2 concentrations in China. Using
data from Ozone Monitoring Instrument (OMI), combined with
socio-economic factors and anthropogenic emission inventories,
4.3.1 Climate change the model provided detailed estimates of NO2 near the ground,
Faced with the major challenges of climate change and its effects subsequently used for environmental quality and epidemic man-
on the planet, several research projects have addressed this issue agement in China.
by proposing novel methods and models for monitoring environ- Moreover, air quality qualification is a key element in governing
mental events. Lou et al. [282] presented a Deep Neural net- and sustainable environmental management. Indeed, Rowley and
work (DNN) using a recursive feature elimination (RFE) algo- Karakuş [288] propose air quality network (AQNet), a deep neural
rithm [293]. The aim of the proposed model is to explore vegeta- network trained on multispectral images from Sentinel-2 satellite,
tion changes using Landsat 5, 8, Sentinel-1 satellites and MODIS merged with NO2 concentrations from Sentinel-5P satellite data.
images, from 1988 to 2018. Over the explored period, the model’s The goal of this research is to form an air quality index from NO2
performance shows an overall accuracy of 83%. Wang et al. [283] concentrations, ozone gases (O3) and particulate matter (PM10),
used a Residual Channel Attention Network (RCAN) [294], it is i.e. airborne particles with an aerodynamic diameter of less than
a model improving image resolution while using attention mech- 10 µm. In addition, Seng et al. [289] exploited an LSTM model
anisms and image fusion, combining details from several images. to predict concentrations of PM2.5, referring also to airborne par-
The aim of this work is to provide a more accurate representation ticles with an aerodynamic diameter of less than 2.5µm, Sulfur
of the vegetation cover from Danjang river basin. Landsat 4, 8 Dioxide (SO2), NO2, O3, and Carbon Monoxide (CO). The input
and MODIS images are used to train the presented model. It is data are ground measurement stations constituting a large spatio-
worth mentioning that the NDVI index [235] is used for the phys- temporal series. The test phase showed outstanding efficiency
ical representation of this vegetation cover. Then, the resulting with an RMSE of 12.48 mg/m3 and an R2 of 91%.
refined representation enabled better reconstruction of the NDVI Besides the area of air quality, quantification of aerosols, i.e.
index for more accurate exploration of change. In the same con- the constituents of fine particles suspended in the atmosphere, is
text, Xin et al. [284] evaluated several methods, in particular neu- essential for studying aerosols impact on the radiation balance,
ral network and random forest, as well as six rule based methods and for understanding the dispersion of pollutants. With this in
to retrieve vegetation phenology using MODIS images and USA mind, Sun et al. [298] dealed with the aerosol absorption prop-
national phenology network data [295]. The results show a big erties using UltraViolet Aerosol Index (UVAI) [299], an index
difference between the different methods in terms of efficiency, reflecting the absorption capacity of aerosols in the ultra-violet
while highlighting the performance of RF model, with an MAE wavelengths. The authors have implemented a DNN using ozone
of 15.16 days. In addition to NDVI, Leaf Area Index (LAI) [240] monitoring instrument (OMI) data for training and ground-based
is a powerful ecological indicator for measuring the interactions aerosol robotic network (AERONET) [300] observations for test,
of flora with the environment, it represents the ratio of the to- a representative map of which is shown in Figure 35, resulting in
tal leaf area of vegetation to the ground surface. Ma and Liang an RMSE of 0.005.

22
 Table 10: Models and proposed methods used for environment monitoring.

Applications Methods and algorithms References

Deep Neural Network and Recursive Feature Elimination method. [282]
Climate change Residual Channel Attention Network. [283]
Rule-based ML, Convolutional Neural Network and Random Forest. [284]
Deep Convolutional Autoencoder. [285]
XGBoost model. [286]
Atmospheric chemistry Random Forest and K-Means clustering algorithm. [287]
Deep Neural Network. [288]
Long Short-Term Memory. [289]
Artificial Neural Network. [16]
Multiscale Geographically Weighted Regression, Random Forest along with SHapley Additive exPlanation
Environmental impacts [290]
model.
Logistic Regression, XGBoost, Light Gradient Boosting Machine, Multi-Layer Perceptron, Long Short-
[291]
Term Memory, Convolutional Long Short-Term Memory and transformers.
DeeplabV3+ architecture. [292]

Index (RSUSEI), a qualitative index of urban ecological condi-

tion [303]. In the same context as Environmental Impact Assess-
ment, Gianquintieri et al. [290] proposed a GeoAI model with a
multi-block architecture in order to assess the impact of agricul-
tural activity on the distribution of PM2.5. Data used are PM2.5
concentrations from Copernicus database and land use maps of
Lombardy region in Italy. The blocks in question are Spearman
rank-based correlation [304] for feature extraction, multiscale ge-
ographically weighted regression (MGWR) [305, 306] for spatial
data enhancement, Random Forest (RF) for identification of areas
with high PM2.5 concentration, and a SHapley Additive exPla-
nation (SHAP) method [307] to compare the impact of several
attributes such as agricultural land use, rate of urbanisation and
presence of water resources. Results of SHAP model showed that
natural areas followed by agricultural land with maize and cereals
have a strong impact on the concentration of PM 2.5.
Figure 35: Example of United States AERONET data dated 06/27/2023 ([301]).
To perform a so-called inverse task, i.e. measuring the environ-
mental effects on agricultural activities, Shevchenko et al. [291]
presented a meta-classifier based on several ML models, specifi-
Finally, the methods and models presented allow for accurate
cally Logistic Regression (LR), CatBoost [308], XGBoost, Light-
models and predict the chemical composition of the atmosphere,
GBM [309], along with various DL models, particularly MLP,
providing a better understanding of atmospheric dynamics and
LSTM, ConvLSTM [310] and a Transformer. The goal of this
processes, assessing air quality and forecasting climatic phenom-
study is to measure the impact of climate change on the suitability
ena.
of agricultural lands in Europe and Asia. The authors use agricul-
tural classification products from Global Food Security-support
4.3.3 Environmental impacts Analysis Data at 1 km resolution (GFSAD1km) [311], together
Given the accuracy and reliability of its techniques, GeoAI is with climatic and morphological data to train and evaluate the
proving extremely valuable in assessing the interactions between model, resulting in an overall accuracy of 86%.
the environment and human activity. Indeed, two concepts are dis- Cao et al. [292] assessed the ecological safety of ice and snow
tinguished, notably Environmental Impact Assessments (EIA), as tourist destinations, particularly in the face of environmental
well as the study of environmental effects on project sites, infras- threats and climate change. To do this, a DeeplabV3+ classifica-
tructures and agricultural land. In this regard, the use of GeoAI is tion model [277] is introduced using Landsat satellite images. As
very useful for monitoring environmental phenomena, measuring shown in Figure 36, the presented model consists of an encoder
the impact of human activities on the integrity of ecosystems, on to extract image characteristics through successive convolutions
physical factors such as air and water quality, on social services and pooling, along with a decoder refining the predictions made
and economic values. by applying convolutions. Then, a factor 4 upsampling and con-
According to Glasson et al. [302], EIA is a systematic approach catenation with the features produced by the decoder are applied,
aiming to analyze and predict the effects of development actions followed by resampling to reduce information loss. The model’s
on the environment. Several AI models contributed to this ex- predictions achieved an overall accuracy of 91%, demonstrating
amination, in fact, an Artificial Neural Network (ANN) is devel- remarkable efficiency compared with traditional methods such as
oped by Al Qadhi et al. [16] to analyse the impact of urbanisa- maximum likelihood [312] and minimum distance method [313].
tion on ecosystems in Abha region of Saudi Arabia. The ANN The models presented in this subsection reflect the state of ex-
in question is trained on Land use and Land Cover (LULC) maps ploration of environmental impact assessment, including both the
using Landsat 4, 5 and 8 images over the years 1990, 2000 and identification of the impacts of urbanization on ecosystems [16],
2010. The proposed ANN is able to obtain an accuracy of 94%, the exploration of the effects of agriculture on environmental pa-
using alongside biophysical, biochemical and biological indica- rameters [290], and the measurement of environmental effects on
tors to develop the Remotely Sensed Urban Surface Ecological agricultural activity [291].

23
The explored examples provide a comprehensive overview of other methods such as SVM, RF and Gradient Boosting Decision
GeoAI’s contribution to the aforementioned theme, not only in Tree. The goal of this work is to predict the response of water
terms of applications, but also with regard to the integration of systems to various environmental conditions. Wegayehu and Mu-
AI models at the heart of the processes in question. The pre- luneh [317] have used data from MODIS instrument to calculate a
sented models have proven their high applicability in environmen- set of environmental indices, including the Normalized Difference
tal monitoring, including the use of geospatial data of all types Water Index (NDWI) [331], i.e. a remotely sensed index used to
to monitor various environmental phenomena, to model environ- identify water surfaces, calculated in the same way as NDVI but
mental impacts and to preserve ecosystems, enabling the engaged using green and near-infrared bands [331]. A set of precipita-
parties to take the necessary measures to protect the environment. tion data are employed, namely IMERG data, CHIRPS precipita-
Following the same logic, the next subsection explores the syn- tion measurements, a case in point is shown in Figure 37, Multi-
ergy between GeoAI and another cutting-edge field, that of water Source Weighted-Ensemble Precipitation Version 2 (MSWEP-
resource management, enabling to respond differently to current V2) to simulate streamflow in three river basins in Ethiopia. Su-
challenges in this area. per ensemble approach, being an advanced approach using mul-
tiple inferences from a multitude of models, called learners, to
4.4 Water resources management and precipita- form a more accurate model [332], is applied. Three models are
exploited in this regard, in particular Weighted Average Super
tion forecasting Ensemble (WASE), using a weighted average of predictions to
Given the vital importance of water resources, a considerable form the result model, Bayesian Model Averaging Super Ensem-
set of research is carried out using innovative GeoAI techniques ble (BMASE), featuring Bayesian inference for the same reason
and methods. This research is aimed at sustainable manage- and Extra Tree Regression Super Ensemble (ETRSE), exploiting
ment of water resources and more targeted interpretation of the Extra Tree Regression (ETR) [333]. This combination resulted in
water-related phenomena, regarding hydrological modelling, wa- an improved accuracy, with an R2 coefficient of up to 77%, and a
ter quality, groundwater management and precipitation. Imple- better adaptation to different scenarios.
menting these techniques enables informed decision-making in
the face of today’s challenges, especially quantifying demand, op-
timizing use and proactively protecting against pollution. Table
11 summarizes the methods and algorithms used.

4.4.1 Hydrological modelling

Thanks to their efficiency and performance, GeoAI methods per-
mit a comprehensive modelling of hydrological events. In this
context, Islam et al. [327] used DL-based space-time image ve-
locimetry (DL-STIV) [328] in order to estimate hydrological pa-
rameters including water flow velocity and flow rate in Asahi
Japanese river. UAV images are collected and used to train the
model. Furthermore, Digital Elevation Models (DEM) derived
from UAV-borne LIDAR data are exploited to validate the results
obtained, i.e. flow velocities with MSEs between 0.01 and 0.24 Figure 37: CHIRPS monthly data example from Morocco and Africa, dated August 2024.
m/s and flow rates estimated with accuracies of up to 11%. For
the purpose of extracting drainage network in low relief areas, Du
et al. [314] have proposed a U-Net architecture [85] using LIDAR 4.4.2 Groundwater
data as input. Results of comparison with Random Forest model Groundwater management is an important area of involvement
show a good performance of the proposed model, obtaining a pre- for researchers in the field of AI. More specifically, and given
cision of 88% and a recall of 89%. its captivating ability to process geospatial data, GeoAI gives wa-
The definition of water-land boundaries provides a better under- ter specialists the opportunity to carry out numerous tasks, such
standing of interactions between terrestrial and aquatic entities, as delineating groundwater recharges, predicting areas containing
giving greater precision in the calculation of flows and the qualifi- potential aquifers, ensuring sustainable management of these re-
cation of environmental impacts. In this context, Sun et al. [315] sources as well as preventive detection of possible contamination.
presented a DL model for coastline extraction. This model is For the purpose of groundwater recharge delineation in the United
based on quadtree decomposition, an algorithm used to divide in- Arab Emirates (UAE), Al-Ruzouq et al. [318] have used a hybrid
put images recursively into four equal regions [329]. This method of convolutional neural network and extreme gradient boosting
is used to reduce computational costs, to eliminate non-interesting (CNN-XGB). For model prediction and training, nine hydrogeo-
areas and to manage scaling. Moreover, the authors use high- logical factors are utilized in this regard, specifically precipitation,
resolution (HR) satellite images from Google Map server using elevation, drainage density, geomorphology, geology, water lev-
the Tile Map Service (TMS). A multi-scale classification is used els, Total Dissolved Solids (TDS), lineament density together with
based on two DL networks, MobileNet V3 [330] and Inception distance to residential areas. The said model is able to achieve
V4 [278]. This model proved a high classification performance, an accuracy of 81%, a recall of 78% and an F1-Score of 80%.
presenting an accuracy improvement of over 6% compared to tra- Along the same lines, Sahour et al. [319] have compared a set of
ditional architectures, for instance Fully Convolutional Network models such as Logistic Regression (LR), SVM and extreme gra-
(FCN) and U-Net [85]. In a different context, Onchan et al. [316] dient boosting (XGB), with the aim of identifying areas of shal-
proposed a Reflectance Transformation - Convolutional Neural low groundwater. Data used in this comparison includes remote
Network (RT-CNN) for generating river bathymetric data from sensing images from Sentinel-1, and Landsat-8 satellites, MODIS
China. The authors use Landsat-7, 8 and Sentinel-2 satellites im- data along with other geomorphological data, in particular eleva-
ages to achieve an improved accuracy of over 18% compared to tion, slope, curvature, distance to undercut features, soil moisture,
24
 Figure 36: DeepLab V3+ flow chart ([292]).

Table 11: Methods and algorithms for water resources management.

Applications Methods and algorithms References

U-Net model. [314]
Quadtree Decomposition. [315]
Hydrological modelling
Reflectance Transformation combined with Convolutional Neural Network. [316]
Weighted Average Super Ensemble, Extra Tree Regression Super Ensemble and Bayesian Model Averaging Super Ensemble. [317]
Convolutional Neural Network - eXtreme Gradient Boosting. [318]
Logistic Regression, Support Vector Machine and eXtreme Gradient Boosting. [319]
Ground water
Long Short-Term Memory and Convolutional Neural Network-Long Short-Term Memory. [320]
Random Forest, Support Vector Regression and Multi-Layer Perceptron. [321]
Ensemble learning approach together with Higher Order Singular Value Decomposition, Deep Neural Networks and Support Vector Ma-
[298]
chine.
Water quality
Transformers, Mixture Density Network, Random Forest, Gradient Boosting and eXtreme Gradient Boosting. [322]
Linear Regression along with Deep Neural Networks. [323]
Heterogeneous Spatiotemporal Attention Fusion Prediction model. [324]
Spatio Temporal Inference Network. [325]
Precipitation forecasting
Fully Convolutional Neural Network, Convolutional Neural Network, Recurrent Neural Network, Long Short-Term Memory, Gated Re-
[20]
current Unit, U-Net and Generative Adversarial Network.
PseudoFlow SpatioTemporal Long Short-Term Memory. [326]

radar backscatter coefficient, NDVI and LST. Test results show an

accuracy of 93% for XGB method, compared to SVM (88%) and
Logistic regression (87%).
In order to predict potential groundwater resources, Nhu et al.
[334] have based on a TensorFlow Deep Neural Network (TF-
DNN), using 12 influencing factors and 733 locations of ground-
water sources. Related factors includes slope, elevation, curva-
ture, NDVI and Normalized Difference Moisture Index (NDMI),
i.e. an index reflecting the water content of vegetation using the
Near InfraRed (NIR) and Short-Wave InfraRed (SWIR) bands
[335]. TF-DNN demonstrated high efficiency, achieving an ac-
curacy of 80% and an F1-Score of 79%. In a similar con-
text, Seo and Lee [320] compared two models, more precisely
the standard LSTM and the combinatorial Convolutional Neu-
ral Network-LSTM (CNN-LSTM) model for predicting spatio- Figure 38: Daily map of TRMM data following the Midwest floods in the USA (06/15/2008).
temporal variations in groundwater storage. Terrestrial water stor-
age anomaly, recorded by GRACE and GRACE-FO satellites,
Another application theme is to improve the sampling frequency
precipitation from Tropical Rainfall Measuring Mission (TRMM)
of groundwater-related data, in which context Sabzehee et al.
satellite, as well as NDVI and modified normalized difference wa-
[321] presented a multitude of models, including RF regression,
ter index (MNDWI) [336] acquired via Landsat 5 and 8 satellites
Support Vector Regression (SVR) [236] and MLP. The aim of this
are employed as input data for the models. Final results have
study is to improve the spatial resolution of groundwater stor-
favored the use of CNN-LSTM in predicting changes in water-
age anomaly, obtained from GRACE sattelite. The model takes
ground storage, acquiring an RMSE of 44.92 mm/month.
as input hydro-climatic variables such as precipitation, LST and
NDVI. RF model have obtained the highest performance with re-
gard to the aim of this study, with an RMSE of 18.36 mm.
From the above, it has emerged that GeoAI plays an important

25
role in forecasting the geographical delineation of groundwater
resources. The results of these forecasts, through the five research
works cited, reflect the performance guaranteed by AI models in
combination with geospatial data, particularly with regard to the
subject in question.

4.4.3 Water quality

Artificial intelligence techniques, combined with geo-
information, can also be a powerful tool for precise quantifi-
cation and qualification of water quality, using remote sensing
products and in-situ measurements, referring to direct field mea-
surements, as useful and relevant acquisition techniques. Sun et
al. [337] proposed an ensemble learning model trained on re-
motely sensed images, based on Integrated Data and Classifier
Fusion via Higher Order Singular Value Decomposition (IDCF-
HOSVD). As indicated, the authors use HOSVD [338], a math-
ematical method for feature extraction and dimension reduction. Figure 39: Water nutrient concentrations from [341].
In addition to generalizing Singular Value Decomposition (SVD)
to multi-dimensional vectors, the classification results of several Optical characteristics, such as diffuse attenuation coefficient and
models, such as neural networks and SVM is applied. The main transparency, indicates precisely the composition of water and the
objective of this work is to measure the parameter ”chlorophyll- presence or absence of various substances. Filisbino Freire da
a” in the water of Lake Nicaragua in Central America, reflecting Silva et al. [342] proposed an SVM architecture combined with a
the degree of nutrimants in the water [339]. The IDCF-HOSVD sigmoı̈de function at the end of model to classify known optical
model showed favourable results when comparing actual results water types (OWTs) and to detect new OWTs. Train, validation
with ground truth measurements. Yang et al. [322] have studied and test data are from Sentinel-2 satellite MultiSpectral Instru-
water quality in krastic wetlands, using multi-spectral and hy- ment (MSI) sensor. Indeed, cross-validation procedure showed
perspectral UAV images, combined with ground measurements. an overall accuracy (OA) of the model of up to 94%.
Concerned process includes the use of various models, namely As a summary, numerous studies have focused on the impact of
transformers, mixture density network [340], RF, XGBoost and GeoAI models in assessing and monitoring water quality. These
gradient boosting. These models are analyzed using SHapley Ad- models accurately predicted a number of parameters related to
ditive exPlanations (SHAP) method [307], to predict the param- the subject, touching on various aspects such as eutrophication,
eters chlorophyll-a (Chla), phycocyanin (PC), as photosynthetic chemical composition and physical characteristics, enabling sus-
pigments, turbidity (Turb), relating to the density of particles sus- tainable monitoring and effective prevention of water pollution.
pended in water, and dissolved oxygen (DO), a significant value
of which represents good water quality. In conclusion, the trans-
4.4.4 Precipitation forecasting
formers achieved the best predictions in terms of PC and DO,
with R2 coefficients of 0.649 and 0.844 respectively, while the AI techniques for water resources management have applications
XGB and gradient boosting models achieved favorable results in in a wide range of fields, including quantification and predic-
terms of chlorophyll-a estimation, with an R2 of 75%. Addition- tion of precipitation. This includes nowcasting, geospatial anal-
ally, Wu et al. [341] exploited remote sensing images to measure ysis of precipitation and prediction of particle natures in rainfall.
water quality in Zhejiang coastal area from China. A Spatiotem- These tasks naturally lead to improved water storage procedures,
poral Deep Belief Network is applied on a dataset of MODIS planned agriculture and optimised water use. In order to carry
images along with field data in order to estimate concentrations out immediate precipitation forecasts, Niu et al. [324] have devel-
of nutrients, especially Dissolved Inorganic Nitrogen (DIN) and oped a Heterogeneous Spatiotemporal Attention Fusion Predic-
Dissolved Inorganic Phosphorus (DIP). The inference results have tion (HST-AFP) network for nowcasting precipitation using input
provided mean values for the coefficients of determination R2 of data such as weather forecasts and radar echo observations. The
48%, it is noted that Figure 39 shows the prediction results over proposed model performed well in comparison to U-Net [85] and
9 years. Besides, Vakili and Amanollahi [323] used Landsat 8 Recurrent Neural Network (RNN) methods, with a Critical Suc-
satellite images to measure water quality through optically in- cess Index (CSI) of 37% in test phase, referring to a weather fore-
active water variables including total Nitrogen (TN) and total casting metric, corresponding to the ratio of correct forecasts to
Phosphorus (TP). Linear Regression and artificial neural network all possible events [343]. With the same objective, Jin et al. [325]
(ANN) are modeled to identify the correlation between Landsat 8 used a set of meteorological data including wind speed, humidity
images as well as TP and TN concentrations. This study showed and temperature. As a basic model, a Spatio Temporal Inference
a high performance of ANN, obtaining an MAE of 5%, an RMSE Network (STIN) is adopted by proving high prediction capabili-
of 6% and an R2 of 86%. ties, several evaluation metrics are used to validate this approach
by obtaining an IOU of 43.15 and 30.37 at the end of the next
hour on two different datasets.
For real-time precipitation forecasting in North America, Sen-
goz et al. [20] employed a multi-model approach using several
neural network architectures, especially Fully Connected Neu-
ral Network (FCNN), CNN, RNN, LSTM, Gated Recurrent Unit
(GRU) [344], U-Net [85] and GAN. The goal of this research is
to improve the results of 8 basic Numerical Weather Prediction
(NWP) [345] models, referring to weather forecasting systems
26
using mathematical models and antecedent meteorological data
[346]. The choice of this approach is made to capture spatial and
temporal characteristics of data, to enhance NWP results and to
reduce its errors. Results in question showed a 17% improve-
ment in MAE values compared to NWP outputs. Furthemore, Luo
et al. [326] proposed PseudoFlow SpatioTemporal Long Short-
Term Memory (PFST-LSTM), in order to refine precipitation fore-
casts over Convolutional LSTM (Conv-LSTM) [310], trajectory
gated recurrent unit (TrajGRU) [347], Spatio-Temporal LSTM
(ST-LSTM) and predictive recurrent neural network (PredRNN)
[348]. The model is a variant of LSTM introducing pseudo-flow
prediction on spatio-temporal data. A 101×101 km spatial echo-
Figure 41: Pair of xBD pre/post earthquake dataset images ([146]).
radar maps from CIKM AnalytiCup 2017 competition, of which a
detailed example dataset is shown in Figure 40, is employed while
demonstrating a better efficiency during evaluation phase with an 4.5.1 Earthquakes
average MSE of 82.11.
New AI approaches using geospatial data enable more effective
earthquake prediction and more accurate quantification of mate-
rial and human damage. Jena et al. [355] used DEM, vector and
attribute data to build a CNN aiming to calculate earthquakes risk
in northern India. This model showed a high degree of effective-
ness in mapping, obtaining an accuracy of 94% and an F1-score
of 91%.
A number of studies use positioning data to identify the iono-
spheric disturbances preceding sedimentation. Xiong et al. [356]
proposed an extended LSTM model with encoder-decoder ar-
Figure 40: Examples of Echo-radar maps of 15 time periods at 4 different altitudes, from
chitecture to detect pre-earthquake ionospheric disturbances, To-
CIKM AnalytiCup 2017 competition ([349]) tal Electron Content-Global Navigation Satellite System (TEC-
. GNSS) measurements are utilized to train and evaluate the pro-
posed model. A comparison is made between the LSTM and
To predict the spatial distribution of stable oxygen isotopes in pre- other models such as the AutoRegressive Integrated Moving Av-
cipitation, Erdélyi et al. [350] have verified several methods, in erage (ARIMA) [369], DNN, RF and SVM, showing a clear
particular Regression Kriging (RK) [351], corresponding to a geo- improvement in the detection of seismic irregularities. Hafeez
statistical method combining Linear Regression (LR) and geospa- et al. [357] used Artificial neural networks (ANNs) to exam-
tial interpolation of a random variable, being the residuals of the ine the atmospheric and ionospheric disturbances preceding Pak-
said LR [351], as well as ML techniques, including standard RF, istan earthquake in 2019. Input data includes Land Surface Tem-
Multiple Regression Random Forest (MRRF) [352] and Regres- perature (LST) records from MODIS instrument, Total Electron
sion Enhanced Random Forest (RERF) [353]. These methods are Content-Global Positioning System (TEC-GPS) and Total Elec-
then validated using MSE metric. Results show varied MSE val- tron Content-Global Ionospheric Map (TEC-GIM) measurements
ues for the RK, MRRF, RERF and RF methods, which are 2.77, before and after the earthquake. Prediction results showed a no-
2.61, 2.99 and 3.08 respectively, demonstrating the effectiveness table efficiency, obtaining temperature variations from 2.8°C to
of ML models compared to the KR model, particularly for MRRF. 13.4°C, as well as TEC-GIM and TEC-GPS value variations of 5
Hence, a wide range of geospatial data are used to forecast pre- TECU for the post-earthquake periods.
cipitation, including satellite images and ground station measure- Thanks to their precision, LIDAR data, photogrammetric and
ments. These data are combined with traditional geostatistical aerial images are a valuable datasource of information for risk
methods, numerical weather models and employed by AI models assessment studies. Torres et al. [358] combined the use of LI-
to obtain accurate predictions validated by ground truth. More- DAR point clouds, aerial orthophotos and satellite images to esti-
over, these AI models are a powerful tool in the service of an- mate the fragility of buildings to earthquakes in Spain. The pro-
other rather critical sector, that of disaster management. This cess comprises three phases: image segmentation, extractions of
field, requiring rapid interventions and accurate forecasting, is buildings and assignment of seismic vulnerability to each foot-
dissected munitiously in the following subsection, while present- print. A multitude of ML methods are exploited, specifically deci-
ing the most interesting research papers in the area. sion trees, SVM, logistic regression and Bayesian networks [370],
which are graphical models formalising the relationships between
4.5 Disaster management various variables to make probabilistic deductions [371]. Test re-
sults provides an F1-Score of 79%, 77%, 80% and 77% for the
The combination of geospatial datasets combined with AI models four models respectively. Zahs et al. [359] developed a change-
in monitoring of natural disasters represents a major advance for- based model to quantify post-earthquake damage. The authors
ward for this type of risk. Despite this, a major limitation exists have used Virtual Laser Scanning (VLS) data, referring to a tech-
in the search and selection of geospatial data relating to natural nique for simulating LIDAR acquisition in a virtual environment
disasters, as these data must be acquired exactly at the time of [372]. This model trains to extract changes from this data, using
the concerned event. Figure 41 shows a concrete example of such k-means clustering and then uses photogrammetric point clouds
data, presenting a pair of pre/post earthquake images taken from for Random Forest classification of damage. As shown in Figure
the xBD dataset [354], constituting an input for related AI models. 42, VLS data is categorized into 4 different classes according to
Furthermore, a number of interesting works are published on the the severity of damage, using a post- and pre-earthquake data pair.
subject, particularly in relation to earthquakes, floods and forest The overall process is validated through a large batch of real point
fires. Table 12 shows the main researches listed in this subsection.
27
 Table 12: Set of methods and algorithms presented for natural disasters management.

Applications Methods and algorithms References

Convolutional Neural Network. [355]
Extended Long Short-Term Memory. [356]
Earthquakes Artificial Neural Network. [357]
Decision Trees, Support Vector Machine, Bayesian Network and Logistic Regression. [358]
K-means and Random Forest. [359]
Artificial Neural Networks, Support Vector Machine, Decision Trees, Bayesian Naive, Random Forest and Fuzzy Logic. [360]
Convolutional Long Short-Term Memory. [361]
Floods
U-Net combined with attention mechanisms. [362]
Fully Convolutional Network-8s. [363]
Artificial Neural Network coupled with a Generative Adversarial Network. [364]
Neural Adaptive Fuzzy Inference System with Artificial Bee Colony optimization algorithm. [365]
Forest fires Logistic regression, MobileNet, Visual Geometry Group, LeNet and Residual Networks. [366]
U-Net architecture. [367]
Convolutional Neural Network. [368]

clouds while achieving an accuracy of 95%. traditional methods, obtaining average precision values of 84%,
recall of 91% and F1-Score of 87%. Montello et al. [376] con-
structed a Multimodal Dataset for Flood delineation (MMFlood)
from Sentinel 1 SAR images, along with elevation data, hydro-
graphic maps and binary annotations. This dataset focuses on the
reliability of SAR data combined with digital elevation models
(DEMs) for flood mapping, particularly in cloudy environments.
A multi-encoder architecture is employed to delineate floods from
this dataset, proving the use of DEMs for greater precision, while
reaching an F1 score of up to 79% for this multimodal approach.
In a similar context, Hashemi-Beni and Gebrehiwot [363] pro-
posed an advanced method to provide more accurate mapping of
flood extents. To this end, UAV and aerial optical images, to-
gether with LIDAR data are collected by the National Oceanic
Figure 42: Example of VLS data post (left) and pre (right) earthquake, each class of data is
represented by an illustrative example ([359]). and Atmospheric Administration (NOAA) [377]. In a first phase,
a particular Fully Convolutional Network (FCN-8s) is applied for
image classification, the application of a tride of 8 for FCN-8s
According to the above-mentioned research, GeoAI contributes
convolution layers shows better performance, proving the choice
effectively to the analysis and prediction of earthquakes, and to
of this architecture. Then, delimitation of flooded areas is im-
the estimation of earthquake damage. The benefits derived are es-
proved by a Regional Growing (RG) segmentation method using
sential for minimizing the effects of seismic events and predicting
DEMs of water levels. The use of data augmentation techniques,
their occurrence.
such as cropping, rotation and translation, resulted in an overall
accuracy of 97%.
4.5.2 Floods Like earthquakes, multiple creative GeoAI models are imple-
Recent advances in GeoAI have enabled to draw up expressive mented to map and forecast floods, essentially employing remote
vulnerability maps and precise models for flood forecasting pur- sensing data, aerial images and LIDAR point clouds in order to
poses. Given this context, Puttinaovarat and Horkaew [360] de- monitor and predict this hyper-descructive phenomenon.
veloped a flood and precipitation forecasting method based on
big geodata, including Web Map Service (WMS), public raster 4.5.3 Forest fires
records, meteorological, hydrological and rowdsourced data. Nu-
Forest fires are one of the most damaging natural disasters, affect-
merous models, specifically SVM, RF, decision trees, fuzzy logic
ing both property and lives of animals and humans. Integrating
[373] and ANNs are evaluated to choose the most performant.
innovative methods into real-time monitoring systems is a very
Results of a k-fold cross-validation [374] favore the use of ANNs,
useful way of identifying vulnerable areas and recognizing the
SVM and RF. Another flood forecasting model is developed by
most critical perimeters in the field. Numerous studies are carried
Moishin et al. [361] using a hybrid ConvLSTM [310] model,
out on the subject of preventing the most susceptible areas to for-
combining the architecture of CNN with LSTM to process spa-
est fires. An ANN is used by Rathod et al. [364], trained by data
tiotemporal data, including rainfall data and Flood Index (IF ),
collected via a drone equipped with the Raspberry Pi4 module and
referring to an index for real-time flood monitoring using daily
coupled with temperature data from a Digital Humidity and Tem-
rainfall measurements [375]. This model has proven its ability to
perature DHT11 sensor. In addition, a GAN is implemented to
handle large-scale spatio-temporal datasets with notable accuracy.
augment the training dataset, while achieving an overall accuracy
In addition to flood forecasting, a number of studies focused on
of 90% for the overall model. Pham et al. [365] aimed to pre-
mapping the affected areas. Li et al. [362] presented a U-Net
dict forest fire susceptibility, through an Adaptive Neural Fuzzy
network [85] coupled with attention mechanisms to better ex-
Inference System (ANFIS) [378]. Several types of data, espe-
tract inundated areas in China. Dual-polarized Sentinel 1 SAR
cially Digital Elevation Models (DEM), NDWI and NDVI indices
images are employed for model training and evaluation. The re-
added to hydrometeorological stations are utilized in this settings.
sults showed a remarkable contribution of this model compared to
Additionally, an Artificial Bee Colony (ABC) algorithm [379] is
28
used to alleviate the problem of data multicollinearity, obtaining the case of COVID-19 as an illustrative example. Table 13 shows
an RMSE of 7.28, an MAE of 5.02 and an R2 of 86% for the all the proposed methods and algorithms.
ANFIS-ABC model.
Delineating forest fires is necessary for reliable and effective an- 4.6.1 Infectious diseases and toxins
ticipation, as mapping the areas affected enables public services to
better understand the severity of damages and to move swiftly on In today’s era of standardized health measures against different
to the post-fire stage. It also enables stakeholders to better quan- types of disease, the use of GeoAI techniques to combat these
tify the impact of these fires on biodiversity and ecological bal- diseases and minimize their spread is proving to be of paramount
ance. Indeed, Chen et al. [366] used Flame2 [380], an improved importance. Venna et al. [385] developed a multi-stage LSTM
version of Flame [381], referring to a customized CNN for forest model to predict influenza in the United States, using data from
fire detection, while comparing it to a combination of DL models, the Centers for Disease Control and Prevention (CDC) [396], il-
in particular LetNet [82], Visual Geometry Group (VGG) [84], lustrated by Figure 44, and Google Flu Trends (GFT) [397], as
MobileNet [382] and ResNet [86]. As illustrated in Figure 43, well as climatic data including precipitation and temperature. The
RGB and infrared (IR) images from drones are used to train com- model’s choice is prouved against Auto-Regression Integrated
parative models. The evaluation stage proved the high detection Moving Average (ARIMA) [369] method. LSTM model achieved
efficiency of FLAME2, reaching an average F1-Score of 83%. the best prediction performance with an RMSE of 1.032 and a
Also, Cho et al. [367] used a U-Net network [85] to map burnt MAPE of 14 over 5 different weekly periods. In a related sce-
areas from PlanetScope’s satellites images (RGB+NIR). Comple- nario, Hu et al. [386] have aimed to predict the real-time spread
mentary NDVI [383] and greyscale texture data are utilized to im- of Inflenza-Like Illness (ILI), using twitter posts and other data
prove the accuracy of obtained results, leading to an F1 score of up from CDC [396], especially average percentages of medical vis-
to 93%. An interesting framework for forest fire classification is its for ILI-like symptoms. An Improved Artificial Tree Back-
developed by Siddique et al. [368], the chosen model in question Propagation Neural Network (IAT-BPNN) is proposed as the base
is a three-layer CNN using fire and non-fire images from Wildfire model, while reaching an MSE of 0.049, an RMSE of 0.038 and
Detection Image Dataset [384]. The proposed framework inte- a MAPE of 0.147. Lu et al. [398] propose AGRONet, a set-
grates Federated Stochastic Gradient Descent (FedSGD) [109] to based approach to estimate the spread of influenza in the United
aggregate this model in the cloud, while obtaining an accuracy of States. The presented model employes a statistical autoregres-
99%. sive approach, used to implement predictions based on historical
trends, Google search frequencies and electronic health records.
A network-based approach is also utilized, using regularized mul-
tivariable regression to take advantage of synchronicities in the
spatio-temporal history. Predictions from 37 states have reached
an RMSE of around 1.2 and a MAPE of around 0.7.

Figure 43: Sample of RGB and IR data used for forest fires detection ([366]).

As mentioned above, several GeoAI models are proposed for for-

est fire prediction and detection, including the use of various data
Figure 44: Weekly U.S map of CDC data, available at the link [399].
collected at different dates, helping to manage the adverse effects
of this disaster on flora and fauna, as well as preserving the lives of
thousands of people. In the same spirit of preservation, healthcare In addition to flu forecasting, Guo et al. [387] compared a multi-
is a field enabling to predict the spread of epidemics and infec- tude of ML models, specifically Support Vector Regression (SVR)
tious diseases, to help analyze and exploit public health data, and [236], step-down linear regression, Gradient Boosted Machine
to provide citizens with the care they need. The following subsec- (GBM) [400], and Least Absolute Shrinkage and Selection Op-
tion surveys the contribution of GeoAI in this field, mentioning erator (LASSO) [401], for predicting dengue fever in Guangdong
the most interesting research in this regard. province from China. Weekly dengue case records and climatic
data including temperature, humidity and precipitation from 2011
to 2014 are used. Comparison results favoured the SVR model,
4.6 Healthcare with an RMSE of 13.52 and an R2 coefficient of 98% during the
Thanks to the reforming effect of artificial intelligence (AI), in- total period.
creased use of geospatial data enables the modelling of spread In addition to infectious diseases, diseases caused by poor eat-
and impact of diseases, the planning of medical campaigns and ing habits and frequent poisoning are also a major focus of study.
the optimizing of resources. These advantages guarantee a “re- Sadilek et al. [388] developed Foodborne IllNess DEtector in
sponsible governance” in front of epidemics and critical health Real time (FINDER), an ML model for automatic detection of
crises. In this section, research projects adressing this issue are foodborne disease sources. Web search data are employed to
listed, while applying GeoAI methods to the surveillance of in- train the model, based on two main components, a web search
fectious diseases such as fevers, influenza and poisoning, together model (WSM) to calculate the probability linking web search and
with public health monitoring, and epidemic modelling, taking food-borne illnesses, besides a Location Model (LM) mapping the

29
 Table 13: Summary of methods and models in Healthcare field.

Applications Methods and algorithms References

Multi-stage Long Short-Term Memory. [385]
Improved Artificial Tree Optimizing Backpropagation Neural Network. [386]
Infectious diseases and toxins
Support vector regression, Gradient Boosted Machine, Least Absolute Shrinkage and Selection Operator, Step-down Linear Regression. [387]
Web search and location models. [388]
High-Resolution Geographical Convolutional Network. [389]
Multi-Layer Perceptron, Support Vector Machine, Linear Regression and Random Forest. [390]
Pandemics Linear Regression. [391]
Convolutional Neural Network combined with an Improved Spatial Transformer and a Multi Grid Module. [392]
Deep neural network and Deep regression model. [393]
Geographically-weighted Gradient Boosting Machine. [394]
Public health
Hybrid Convolution Neural Network. [395]

searches identified and analyzed by the WSM. Identified restau-

rants are found to be three times more likely to be hazardous than
those identified by traditional methods, i.e. periodic inspections
and those based on user complaints.
In summary, clinical observations and navigational research tak-
ing into account geographical locations are exploited to predict
trends of infectious diseases and toxins, in order to speed up the
response of the relevant authorities in the event of any crises. The
next subsection looks at the case of pandemics, more prone to
spread and therefore more pernicious.

4.6.2 Pandemics : COVID-19

GeoAI’s analysis capacity have enabled interested parties to ac-
curately model epidemics of various types, and to monitor their
evolution over time and space. More specifically, and given the
major impact of the COVID-19 pandemic on the whole world, a
Figure 45: Architecture of ECO4CO ([391]).
great deal of work in recent years focused on the contribution of
GeoAI to predic the impact of this pandemic. In fact, it is decided
to concentrate mainly on work relating to this pandemic, since it For monitoring the so-called pandemic, Zhang et al. [392] pre-
has attracted the greatest interest. sented a multi-module DL approach, consisting of a CNN with
Indeed, Zhao et al. [389] proposed COVID-19 high-resolution ge- batch normalization, an improved spatial transformer and a multi-
ographical convolutional network (CHGCN), a novel model for grid module in order to identify individuals using video surveil-
predicting the spread of human-to-human respiratory infectious lance. As illustrated in Figure 46, this approach is implemented
disease (H-HRID). An innovative feature of the model incorpo- via a 3D visualisation using the city engine tool, providing a better
rates the notion of spatial neighborhood through attention mecha- understanding of the proximity and interactions between individu-
nisms, resulted in an average accuracy of 75%. In the same logic, als. It is noted that this model achieved a high detection efficiency,
Da silva et al. [390] carried out a spatio-temporal predictive re- while obtaining a mAP of 78%.
search of COVID-19 transmission in Brazil. Various models are
used in this analysis, namely MLP, Linear Regression (LR), SVM
and RF. In conclusion, MLP and LR give the best prediction effi-
ciency, obtaining RMSE values of 4% and 2% respectively.
Another field of application is the monitoring of COVID-19 pan-
demic clusters in order to ensure an in-depth understanding of
transmissions. Atek et al. [391] present Earth Cognitive System
for COVID-19 (ECO4CO), a comprehensive system for manag-
ing COVID-19. As shown in Figure 45, ECO4CO is utilized by
two groups of users, end users who complete the monitoring and
control tasks, and admins who configure these tasks and manage
input data. Images from Pleiades satellite are used to train a DL
network for identifying gatherings of people and vehicles, while
giving reliable results.
Figure 46: Sample 3D model generated in [392] for individuals proximity.

Furthermore, Tan et al. [393] studied the effect of COVID-19 on

the concentration of PM2.5 in China. The authors use Aerosol
Optical Depth (AOD) [281] from MODIS along with UltraViolet
Aerosol Index (UVAI) [299] from Sentinel 5P sattelite. The ap-
proach consists of applying a DNN to correlate image data with
PM2.5 concentrations and a Deep Regression model for predic-
30
tions. This study have demonstrated high predictive capability, Ultimately, numerous works have used original GeoAI models for
resulting in an MAE of around 17 µg/m3 and an R2 of 53%. public health, explored via remote sensing images and ground
From this census, it is clear that GeoAI constitutes a key tool in the measurement data, and exploited for monitoring citizens behav-
management of pandemics. Taking COVID-19 as an illustrative ior, for predicting pollution measurements and for analyzing their
example, GeoAI enables real-time monitoring of of this pandemic correlation with human activities. Indeed, implementing these
while tracking its impact on the population. Spatio-temporal data models in decision-making systems is mandatory to ensure more
and updated statistics are used to this end, providing unrivalled effective governance in the face of today’s challenges.
support in the fight against and control of the pandemic, opti-
mizing the human and material resources allocated for this pur-
pose. Dealing with this subject, closely linked to the health and 5 Discussions and outlooks
the safety of communities, obviously leads to take a particular in-
terest in a broader challenge, that of public health. 5.1 Results
Thanks to the review carried out, a body of research focusing on
4.6.3 Public health GeoAI is identified, while integrating the spatial dimension into
In this subsection, the use of GeoAI models in public health is the algorithms developed in this context. As shown in Figure 48,
explored, with a few examples mainly concerning mobile health the majority of the research papers cited are fairly recent, i.e. 2022
and air quality, as a parameter measuring the pollution of urban and beyond.
environments. According to Kahn et al. [402], mobile health con-
cerns the use of wireless communication devices for public health
and clinical practice. Based on this, Kumar et al. [403] presented
the Center of Excellence for Mobile Sensor Data-to-Knowledge
(MD2K), effectively created in October 2014 and funded by the
National Institutes of Health in the United States, to the end of in-
terpreting a large mass of data from portable devices, mainly GPS
location data and user behaviour. In addition, creative IA models
such as puffMarker, mCrave and cStress are employed to train this
data. These models show significant potential for addressing pub-
lic health issues and ensuring rapid, accurate medical intervention
when needed.
Pollution is one of the main problems affecting the health of in-
dividuals and the purity of the environment, making it one of
the main safety factors guaranteeing public and collective health. Figure 48: Yearly papers numbers.
This is why many researchers are focused on this issue and its
impact on collective health. Indeed, Zhan et al. [394] have im- Figure 49 presents a cloud of the most frequently used key-
plemented Geographically-Weighted Gradient Boosting Machine words in surveyed papers, offering a global perspective on the
(GW-GBM) to accurately predict PM2.5 levels in China. The themes adressed. Terms such as learning, data, image, neural
principle of this approach is to integrate geographical weights, network, detection, regression and classification are used fre-
i.e. spatial smoothing kernels, into a GBM model [400]. Aerosol quently, proving the interest of this review for learning methods,
data from MODIS instrument, ground-based PM2.5 measure- especially neural networks. In addition, several categories of data
ments, demographic and climatological records are used to feed are employed to serve multiple tasks, particularly detection, clas-
the model, achieving an RMSE of 23.0 µg/m3 and an R2 coeffi- sification and regression. Without forgetting to emphasize specific
cient of 76%. In addition, Di et al. [395] presented a hybrid CNN methods such as spatio-temporal series, feature extraction and
to predict PM2.5 concentrations in the United States. Trained attention calculation.
and tested datasets comprises Aerosol Optical Depth (AOD) im-
ages from MODIS instrument, Absorbing Aerosol Index (AAI)
data from OMI instrument, illustrated in Figure 47 as an exam-
ple, along with meteorological data and Land-Use terms. The
deployed CNN performed very well, achieving an average coeffi-
cient R2 of 84%.

Figure 49: Most used keywords.

Figure 47: Example of AAI data from OMI instrument, dated 08/11/2024 ([404]).

31
Besides, Figure 50 shows the number of papers cited for each re- • More reliable results: the complexity of used algorithms, the
search area, while demonstrating, in general, an equilibrium bal- consistency of data and the ability to control results all mean
ance in the interest accorded to each chosen application. Despite improved accuracy and reduced operating costs.
this, a relative attention is given to the fields of urban planning
and water resource management. • Interoperability: data standardisation and continuous learning
processes ensure the compatibility and consistency of geospa-
tial methods.
• Performing tasks using new approaches that are totally differ-
ent from those adopted by conventional methods, while ensur-
ing the integration of spatial dimension into the analyses imple-
mented.
• Trend forecasting: analysis of dynamics and changes to better
predict events and facts.
• Analysis-based decision-making: GeoAI facilitates strategic
decision-making based on big geodata, enabling on-the-spot in-
teraction with alerts and emergencies.

5.3 Fundamental aspects

Several key aspects emerge from the discussion, notably the rel-
evance of the methods discussed in this paper for exploiting
Figure 50: Number of papers by field application.
geospatial data even with their complexity, in order to implement
high-performance AI models despite, in many cases, the unavail-
Taking this analysis a step further, Figure 52 illustrates the number ability of input data. Furthermore, the versatility of GeoAI mod-
of papers for each chosen research sub-direction. Furthermore, els makes this field broader than ever, particularly in front of the
the papers highlighting each GeoAI application area are balanced global challenges facing organizations and individuals. In spite
in terms of coverage, indicating the maturity of the axes selected. of these challenges, the integration of high-tech AI in the service
Moreover, the map in Figure 51 display the geographical distri- of geospatial data leads to unavoidable results. In fact, this anal-
bution of all selected papers according to the university of the ysis lays the foundations for a complete answer to RQ2 and 3,
corresponding author, This map clearly demonstrates the global addressed in the next two subsections.
coverage of this study, citing research works from all five conti-
nents. Moreover, Chinese universities dominate this exploration,
followed by the American universities. 5.4 GeoAI challenges (RQ2)
Based on the foregoing, a number of challenges are identified.
5.2 Discussions about RQ1 First of all, data heterogeneity poses enormous problems, es-
pecially when dealing with complex problems requiring several
In Section 4, a range of GeoAI applications are explored, through types of data, often unstructured. This issue being highly exposed
interesting research in areas ranging from Precision agriculture to in the case of data from multiple sources. Additionally, and given
healthcare. Starting with the first area, innovative AI models are the volume and variety of geospatial data, managing and exploit-
revolutionizing the agricultural sector, including models for agri- ing big geodata remains a significant problem, despite advances
cultural crop mapping, yield prediction and creative methods for in research in this field, notably [289, 360, 361].
precision irrigation in order to optimize the agricultural produc- Moreover, there is an exponential need to obtain increasingly ac-
tivity. curate models, especially when dealing with simulated or a lim-
A very interesting field of attack is explored next, that of urban ited volume of data. Also, data quality control remains a subject
planning, logistics and transport, covering information extraction, for discussion, in particular with regard to open-access data, given
urban dynamics, smart cities, transportation, green infrastructure the need to verify its veracity and correct it where necessary. Fur-
management and biodiversity. The models involved enabled auto- thermore, geospatial data sovereignty is one of the most frequent
mated monitoring and highly optimized urban development. Af- problems, as this data is more vulnerable to cyber-attacks, given
terwards, the topic of environmental management is discussed. its richness and implications. Besides, real-time data processing
Then, GeoAI is widely used to assess climate change, quantify requires careful thought in terms of training, updating and feeding
atmospheric chemistry and measure the impact of human activity in newly introduced data.
on environment components, especially fauna, flora and abiotic In contrast, numerous challenges related to the applicative aspect
environments. of GeoAI are mentioned. First of all, optimizing the training and
Subsequently, research linking water resource management to inference times of models is a key objective, particularly in view
GeoAI is covered in this work, with a focus on hydrological mod- of geospatial data complexity, requiring large amounts of mem-
elling, groundwater monitoring and water quality. Disasters man- ory and more optimized hardware architectures. The particular
agement is addressed in a set of research papers, including earth- aspect of geospatiality leads to explore another confrontation, ex-
quakes, floods and forest fires. plicitly the development of specific model architectures for pro-
The healthcare is also explored, with a list of research projects, cessing geospatial data, especially regarding elementary tasks as
covering forecasting of infectious diseases using geospatial data feature extraction and score calculation for attention mechanisms.
and methods, measures to strengthen public health as well as pan- As well, the development of models capable of handling updated
demic control, highlighting COVID-19 as a major example. geospatial data is an urgent need, particularly in relation to re-
Altogether, this review gives a clear understanding of the contri- mote sensing data, improved each time the sattelite is emptied.
bution of GeoAI in different fields of application, guaranteeing:
32
 Figure 51: Number of papers by country.

This need can be met by a periodic fine-tuning, improving the ac- Geospatial Consortium (OGC) [419], constitutes an interesting
curacy of these models over time. outlook. The aim of defining such standards is to ensure the qual-
Security of GeoAI models against adversarial attacks requires ity of the methods implemented and facilitate the integration and
specific adaptations so as to cope with the sensitivity and vul- accessibility of models. Moreover, interpreting geospatial data,
nerability of data. Although a few research works address this particularly remotely sensed images, is an affordable target area
topic including [405, 406, 407], progress in terms of scientific lit- by coupling computer vision and Natural Language Processing
erature is insufficiently explored so far in this context. Another (NLP) models [420]. The aim of this combination is to replace
challenge worth mentioning is scalability, referring to the ability the enormous efforts of specialists while completing the required
of a model to handle a continuous growth in the number of tasks tasks in a relatively short time. The work of [421, 422] combining
or data to be processed. This matter is due to the great evolution NLP models, more precisely Large Language Model (LLM) with
of geospatial sector in recent years, whether in terms of data ac- CNNs, represent a good start in spite of the limitation of results.
cessibility and the growing number of requests. Moreover, many In closing, it is worth mentioning that GeoAI realizes significant
situations require the use of multiple GeoAI models in a single progress, not only in terms of the evolution of methods and al-
task. This configuration brings a multitude of benefits in terms of gorithms, but also in terms of the broad scope of coverage of the
results and interoperability, but also presents many issues, often themes in question. The diversity of GeoAI’s applications demon-
related to model validation and implementation costs, necessitat- strates the breadth of its advances, a breadth extending to other
ing the implementation of distributed architectures [408], while areas of geospatial data analysis and processing. For example, a
guaranteeing modularity of deployment, secure data sharing, etc. wealth of research cited demonstrates the importance of AI-GIS
synergy, notably [392, 254].
5.5 Perspectives and conclusions (RQ3) As far as the application side of GeoAI is concerned, an attractive
societal and environmental impact is emerging, underlining the
Faced with the challenges described above, new research horizons potential offered by GeoAI to help solve problems of different
need to be opened up, notably the conceptualisation of new model dimensions. In this respect, GeoAI provides solutions insofar as
architectures specifically adapted to the nature of geospatial data. novel models are proposed to optimize the use of natural resources
As well as the integration of AI explainability in geospatial data and agricultural land, ensure state-of-the-art urban planning with
processing. It is noted that little research works used this concept certainty. Also, GeoAI solemnly undertakes to respect social in-
for geospatial tasks, such as [409, 410, 411, 412]. Indeed, making tegrity in terms of availability of commercial services, protection
model behavior more comprehensible and plausible with regard against natural disasters and healthcare, etc.
to the standards in question remains a promising prospect. A crucial point to raise is the need to guarantee a professional
According to van Kranenburg [413], Internet of Things (IoT) framework for collaboration between academics and profession-
refers to an infrastructure of physical sensors networks, or virtual als. This collaboration manifests itself into a mutual and effective
objects based on standard communication protocols, to exchange transfer of knowledge, and an enhancement of GeoAI’s learning
data between themselves, or with other external systems. Fur- methods through the inclusion of a practical dimension. This
thermore, the use of IoT devices to collect geospatial data and interdisciplinary approach irrevocably certifies irreproachable re-
then employ it in AI models is an existing technique, as shown sults by offering unexpected solutions to the most intractable
by [414, 368, 415, 416], but it is not sufficiently explored to take problems.
advantage of the efficiency, automation and optimization offered Hence, it is concluded that GeoAI has a promising future. Given
by IoT. the exponential progress of AI, it is difficult to predict the next
In addition, the definition of standards for the training and in- degree of evolution related to this theme. Furthermore, this
ference of GeoAI models, similar to those governing the access work will certainly help researchers to better understand the
and use of geospatial data, such as Spatio-Temporal Asset Cat- current state of GeoAI, as well as the problems it poses, in
alog (STAC) [417], Geographically Encoded Objects for Really order to continue making progress in such a hopeful subject.
Simple Syndication (GeoRSS) [418], and the standards of Open
33
 Figure 52: Detailed statistics.

Declaration of Competing Interest [4] Lijia Chen, Pingping Chen, and Zhijian Lin. “Artificial
intelligence in education: A review”. In: Ieee Access 8
The writing team of this paper mentions that no financial or (2020), pp. 75264–75278. DOI: 10 . 1109 / ACCESS .
lucrative interest is behind this work. 2020.2988510.
[5] Ransome Epie Bawack et al. “Artificial intelligence in E-
Aknowledgments Commerce: a bibliometric study and literature review”.
In: Electronic markets 32.1 (2022), pp. 297–338. DOI:
The authors would like to thank the Artificial Intelligence Geo- 10.1007/s12525-022-00537-z.
decision Networking Optimisation And Cybersecurity (AGNOX)
research team at the National Institute of Posts and Telecommu- [6] Tarik Mitran, Ram Swaroop Meena, and Abhishek
nications (INPT) - Morocco for their unconditional support. Chakraborty. “Geospatial technologies for crops and
soils: an overview”. In: Geospatial technologies for crops
and soils (2021), pp. 1–48. DOI: 10 . 1007 / 978 - 981 -
References 15-6864-0_1.
[7] KV Raju, VR Hegde, and Satish A Hegde. Geospatial
[1] Caiming Zhang and Yang Lu. “Study on artificial intel- Technologies for Agriculture: Case Studies from India.
ligence: The state of the art and future prospects”. In: Springer, 2019. DOI: 10.1007/978-3-319-96646-5.
Journal of Industrial Information Integration 23 (2021),
p. 100224. DOI: 10.1016/j.jii.2021.100224. [8] Praveen Kumar Rai, Varun Narayan Mishra, Prafull
Singh, et al. Geospatial Technology for Landscape and
[2] Fei Jiang et al. “Artificial intelligence in healthcare: past, Environmental Management. Springer, 2022. DOI: 10 .
present and future”. In: Stroke and vascular neurology 2.4 1007/978-981-16-7373-3.
(2017). DOI: 10.1136/svn-2017-000101.
[9] Olga Butenko, Stanislav Horelik, and Oleh Zynyuk.
[3] Ashkan Entezari et al. “Artificial intelligence and ma- “Geospatial data processing characteristics for environ-
chine learning in energy systems: A bibliographic mental monitoring tasks”. In: Architecture, Civil Engi-
perspective”. In: Energy Strategy Reviews 45 (2023), neering, Environment 13.1 (2020), pp. 103–114. DOI: 10.
p. 101017. DOI: 10.1016/j.esr.2022.101017. 21307/acee-2020-008.

34
[10] Daniel Z Sui. “Geospatial technologies and homeland se- [24] Helen Couclelis. “Artificial intelligence in geography:
curity: An overview”. In: Geospatial Technologies and Conjectures on the shape of things to come”. In: The pro-
Homeland Security: Research Frontiers and Future Chal- fessional geographer 38.1 (1986), pp. 1–11. DOI: 10 .
lenges (2008), pp. 1–9. DOI: 10.1007/978- 1- 4020- 1111/j.0033-0124.1986.00001.x.
8507-9_1. [25] Stan Openshaw and Christine Openshaw. Artificial intel-
[11] Zhenjiang Shen. Geospatial techniques in urban plan- ligence in geography. John Wiley and Sons, Inc., 1997.
ning. Springer Science & Business Media, 2012. DOI: 10. ISBN : 978-0-471-96991-4.
1007/978-3-642-13559-0. [26] Huina Mao et al. “Geoai 2017 workshop report: the 1st
[12] GP Obi Reddy. Geospatial Technologies in Land acm sigspatial international workshop on geoai:@ ai and
Resources Mapping, Monitoring and Management. deep learning for geographic knowledge discovery: Re-
Springer, 2018. DOI: 10.1007/978-3-319-78711-4. dondo beach, ca, usa-november 7, 2016”. In: SIGSpatial
[13] Bin Jiang and Xiaobai Yao. “Geospatial analysis and Special 9.3 (2018), pp. 25–25. DOI: 10.1145/3178392.
modeling of urban structure and dynamics: An overview”. 3178408.
In: Geospatial analysis and modelling of urban structure [27] Yingjie Hu et al. “GeoAI at ACM SIGSPATIAL: progress,
and dynamics (2010), pp. 3–11. DOI: 10.1007/978-90- challenges, and future directions”. In: Sigspatial Spe-
481-8572-6_1. cial 11.2 (2019), pp. 5–15. DOI: 10 . 1145 / 3377000 .
[14] Gouri Sankar Bhunia and Pravat Kumar Shit. Geospatial 3377002.
analysis of public health. Springer, 2019. DOI: 10.1007/ [28] Song Gao. Geospatial artificial intelligence (GeoAI).
978-3-030-01680-7. Vol. 10. Oxford University Press New York, 2021. DOI:
[15] Fazlay S Faruque. “Geospatial Technology for Human 10.1093/OBO/9780199874002-0228.
Well-Being and Health: An Overview”. In: Geospatial [29] Wenwen Li and Chia-Yu Hsu. “GeoAI for large-scale im-
Technology for Human Well-Being and Health (2022), age analysis and machine vision: recent progress of arti-
pp. 1–27. DOI: 10.1007/978-3-030-71377-5_1. ficial intelligence in geography”. In: ISPRS International
[16] Saeed Alqadhi et al. “Applying deep learning to manage Journal of Geo-Information 11.7 (2022), p. 385. DOI: 10.
urban ecosystems in arid Abha, Saudi Arabia: Remote 3390/ijgi11070385.
sensing-based modelling for ecological condition assess- [30] Bala Bhavya Kausika et al. “GeoAI for detection of solar
ment and decision-making”. In: Heliyon 10.4 (2024). DOI: photovoltaic installations in the Netherlands”. In: Energy
10.1016/j.heliyon.2024.e25731. and AI 6 (2021), p. 100111. DOI: 10.1016/j.egyai.
[17] Ignazio Gallo et al. “In-season and dynamic crop map- 2021.100111.
ping using 3D convolution neural networks and sentinel- [31] GeoInfotech. Understanding The Difference Between
2 time series”. In: ISPRS Journal of Photogrammetry and GIS and Geospatial data. Accessed: October 03, 2024.
Remote Sensing 195 (2023), pp. 335–352. DOI: 10.1016/ July 31, 2022. URL: https://geoinfotech.ng/what-
j.isprsjprs.2022.12.005. is - the - difference - between - geospatial - and -
[18] Thomas Moranduzzo and Farid Melgani. “A SIFT-SVM gis/ (visited on 01/02/2024).
method for detecting cars in UAV images”. In: 2012 IEEE [32] Xiran Zhou. GeoAI-Enhanced Techniques to Support Ge-
International Geoscience and Remote Sensing Sympo- ographical Knowledge Discovery from Big Geospatial
sium. Ed. by IEEE IGARSS. IEEE. 2012, pp. 6868–6871. Data. Tech. rep. Arizona State University, 2019.
DOI : 10.1109/IGARSS.2012.6352585.
[33] scopus. Scopus. Accessed: October 03, 2024. 2024. URL:
[19] Ekaterina Kalinicheva et al. “Unsupervised change de- https://www.scopus.com.
tection analysis in satellite image time series using deep [34] Google. Google Scholar. Accessed: June 13, 2024. 2024.
learning combined with graph-based approaches”. In: URL : https://scholar.google.com.
IEEE Journal of Selected Topics in Applied Earth Obser-
vations and Remote Sensing 13 (2020), pp. 1450–1466. [35] ScienceDirect. Machine learning application in GIS and
DOI : 10.1109/JSTARS.2020.2982631. remote sensing: An overview. Accessed: October 03,
2024. 2023. URL: https : / / www . sciencedirect .
[20] Cenker Sengoz et al. “Machine learning approaches to im- com/.
prove north american precipitation forecasts”. In: IEEE
Access (2023). DOI: 10.1109/ACCESS.2023.3309054. [36] IEEE Xplore. Machine learning application in GIS and
remote sensing: An overview. Accessed: October 03,
[21] Roderick Lammers et al. “Prediction models for urban 2024. 2023. URL: https://ieeexplore.ieee.org/.
flood evolution for satellite remote sensing”. In: Journal
of Hydrology 603 (2021), p. 127175. DOI: 10.1016/j. [37] Springer. SpringerLink. Accessed: October 03, 2024.
jhydrol.2021.127175. 2024. URL: https://link.springer.com/.
[22] Terence R Smith. “Artificial intelligence and its applica- [38] MDPI. MDPI Platform. Accessed: October 03, 2024.
bility to geographical problem solving”. In: The Profes- 2023. URL: https://www.mdpi.com/.
sional Geographer 36.2 (1984), pp. 147–158. DOI: 10 . [39] Yongjun Xu et al. “Artificial intelligence: A powerful
1111/j.0033-0124.1984.00147.x. paradigm for scientific research”. In: The Innovation 2.4
[23] John E Estes, Charlene Sailer, and Larry R Tinney. “Ap- (2021). DOI: 10.1016/j.xinn.2021.100179.
plications of artificial intelligence techniques to remote [40] KC Santosh, Nibaran Das, and Swarnendu Ghosh. Deep
sensing”. In: The Professional Geographer 38.2 (1986), learning models for medical imaging. Academic Press,
pp. 133–141. DOI: 10 . 1111 / j . 0033 - 0124 . 1986 . 2021.
00133.x.

35
[41] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. [55] Song Gao et al. “Special issue on geospatial artificial in-
“Deep learning”. In: nature 521.7553 (2015), pp. 436– telligence”. In: GeoInformatica 27.2 (2023), pp. 133–136.
444. DOI: 10.1038/nature14539. DOI : 10.1007/s10707-023-00493-6.
[42] Bradley C Love. “Comparing supervised and unsuper- [56] Nishi Srivastava and Nisheeth Saxena. “Applications of
vised category learning”. In: Psychonomic bulletin & Artificial Intelligence and Machine Learning in Geospa-
review 9.4 (2002), pp. 829–835. DOI: 10 . 3758 / tial Data”. In: Emerging Trends, Techniques, and Appli-
BF03196342. cations in Geospatial Data Science. Ed. by John Smith
[43] Pádraig Cunningham, Matthieu Cord, and Sarah Jane De- and Jane Doe. IGI Global, 2023, pp. 196–219. DOI: 10.
lany. “Supervised learning”. In: Machine learning tech- 4018/978-1-6684-7319-1.ch010.
niques for multimedia: case studies on organization and [57] Wenwen Li. “GeoAI and Deep Learning”. In: The Inter-
retrieval. Springer, 2008, pp. 21–49. DOI: 10.1007/978- national Encyclopedia of Geography (2021).
3-540-75171-7_2. [58] Krzysztof Janowicz. “Philosophical foundations of geoai:
[44] Gulden Kaya Uyanik and Necse Guler. “A study on mul- Exploring sustainability, diversity, and bias in geoai and
tiple linear regression analysis”. In: Procedia-Social and spatial data science”. In: Handbook of Geospatial Arti-
Behavioral Sciences 106 (2013), pp. 234–240. DOI: 10. ficial Intelligence. CRC Press, 2023, pp. 26–42. ISBN:
1016/j.sbspro.2013.12.027. 9781003308423.
[45] Yingjie Hu et al. “GeoAI at ACM SIGSPATIAL: progress, [59] Gloria Bordogna and Cristiano Fugazza. “Artificial In-
challenges, and future directions”. In: Sigspatial Spe- telligence for Multisource Geospatial Information”. In:
cial 11.2 (2019), pp. 5–15. DOI: 10 . 1145 / 3377000 . ISPRS International Journal of Geo-Information 12.1
3377002. (2022). Accessed: October 03, 2024, p. 10. DOI: 10 .
[46] Yingjie Hu et al. “A five-year milestone: reflections on 3390/ijgi12010010.
advances and limitations in GeoAI research”. In: Annals [60] Shan Liu et al. “Geospatial AI in Earth Observation,
of GIS 30.1 (2024), pp. 1–14. DOI: 10.1080/19475683. Remote Sensing, and GIScience”. In: Applied Sciences
2024.2309866. 13.22 (2023). Accessed: October 03, 2024, p. 12203. DOI:
[47] Lokendra PS Chauhan and Shashi Shekhar. “GeoAI– 10.3390/app132212203.
accelerating a virtuous cycle between AI and Geo”. In: [61] Shyama Mohan and MVSS Giridhar. “A brief review of
Proceedings of the 2021 Thirteenth International Con- recent developments in the integration of deep learning
ference on Contemporary Computing. Ed. by John Doe with GIS”. In: Geomatics and Environmental Engineer-
and Jane Smith. 2021, pp. 355–370. DOI: 10 . 1145 / ing 16.2 (2022), pp. 21–38. DOI: 10.7494/geom.2022.
3474124.347417. 16.2.21.
[48] Roberto Pierdicca and Marina Paolanti. “GeoAI: a re- [62] Chikodinaka Vanessa Ekeanyanwu et al. “Merging GIS
view of artificial intelligence approaches for the inter- and Machine Learning Techniques: A Paper Review”. In:
pretation of complex geomatics data”. In: Geoscientific Journal of Geoscience and Environment Protection 10.9
Instrumentation, Methods and Data Systems Discussions (2022), pp. 61–83. DOI: 10.4236/gep.2022.109004.
2022 (2022), pp. 1–35. DOI: 10 . 5194 / gi - 11 - 195 - [63] David R Cox. “The regression analysis of binary se-
2022,2022. quences”. In: Journal of the Royal Statistical Society Se-
[49] Gengchen Mai et al. “On the opportunities and challenges ries B: Statistical Methodology 20.2 (1958), pp. 215–232.
of foundation models for geoai (vision paper)”. In: ACM DOI : 10.1111/j.2517-6161.1958.tb00292.x.
Transactions on Spatial Algorithms and Systems (2024). [64] J. Ross Quinlan. “Induction of decision trees”. In: Ma-
DOI : 10.1145/3653070. chine learning 1 (1986), pp. 81–106. DOI: 10 . 1007 /
[50] Renee Sieber. “GeoAI and its implications”. In: Inter- BF00116251.
national Encyclopedia of Geography: People, the Earth, [65] Leo Breiman. “Random forests”. In: Machine learning 45
Environment and Technology (2016), pp. 1–8. DOI: 10 . (2001), pp. 5–32. DOI: 10.1007/978- 1- 4419- 9326-
1002/9781118786352.wbieg2043. 7_5.
[51] Wenwen Li. “GeoAI: Where machine learning and big [66] Hossein Sahour et al. “Random forest and extreme gra-
data converge in GIScience”. In: Journal of Spatial In- dient boosting algorithms for streamflow modeling using
formation Science 20 (2020), pp. 71–77. DOI: 0.5311/ vessel features and tree-rings”. In: Environmental Earth
JOSIS.2020.20.658. Sciences 80 (Nov. 2021). DOI: 10.1007/s12665-021-
[52] Li Chen and Wei Zhang. “Exemplification on Potential 10054-5.
Applications and Scenarios for GeoAI”. In: Proceedings [67] Corinna Cortes and Vladimir Vapnik. “Support-vector
of the IEEE International Conference on Artificial Intelli- networks”. In: Machine learning 20 (1995), pp. 273–297.
gence. Ed. by IEEE International Conference on Artificial [68] Eddy Mayoraz and Ethem Alpaydin. “Support vector ma-
Intelligence. IEEE, 2023, pp. 345–356. DOI: 10.1109/ chines for multi-class classification”. In: International
ACEDPI58926.2023.00054. Work-Conference on Artificial Neural Networks. Ed. by
[53] Krzysztof Janowicz et al. GeoAI: spatially explicit arti- International Work-Conference on Artificial Neural Net-
ficial intelligence techniques for geographic knowledge works. Springer. 1999, pp. 833–842. DOI: 10 . 1007 /
discovery and beyond. Accessed: October 03, 2024. 2020. BFb0100551.
[54] Song Gao. “A review of recent researches and reflections [69] Makarand Velankar. “Study of Emotion Detection in
on geospatial artificial intelligence”. In: Geomatics and Tunes Using Machine Learning”. In: International Jour-
Information Science of Wuhan University 45.12 (2020), nal of Scientific and Research Publications 5 (Nov. 2015),
pp. 1865–1874. DOI: 10.13203/j.whugis20200597. p. 441.
36
[70] Evelyn Fix. Discriminatory analysis: nonparametric dis- [85] Olaf Ronneberger, Philipp Fischer, and Thomas Brox.
crimination, consistency properties. Vol. 1. USAF school “U-Net: Convolutional Networks for Biomedical Im-
of Aviation Medicine, 1985. age Segmentation”. In: Medical Image Computing and
[71] Leo Breiman. “Bagging predictors”. In: Machine learning Computer-Assisted Intervention – MICCAI 2015. Ed. by
24 (1996), pp. 123–140. DOI: 10.1007/BF00058655. Springer. Springer, 2015, pp. 234–241. DOI: 10.1007/
978-3-319-24574-4_28.
[72] Yoav Freund and Robert E Schapire. “A decision-
theoretic generalization of on-line learning and an appli- [86] Kaiming He et al. “Deep Residual Learning for Image
cation to boosting”. In: Journal of computer and system Recognition”. In: Proceedings of the IEEE Conference on
sciences 55.1 (1997), pp. 119–139. DOI: 10.1006/jcss. Computer Vision and Pattern Recognition (CVPR). Ed. by
1997.1504. IEEE. IEEE, 2016, pp. 770–778. DOI: 10.1109/CVPR.
2016.90.
[73] David H Wolpert. “Stacked generalization”. In: Neural
networks 5.2 (1992), pp. 241–259. DOI: 10 . 1016 / [87] J Redmon. “You only look once: Unified, real-time ob-
S0893-6080(05)80023-1. ject detection”. In: Proceedings of the IEEE conference
on computer vision and pattern recognition. Ed. by IEEE
[74] T Chen. “Xgboost: extreme gradient boosting”. In: R conference on computer vision and pattern recognition.
package version 0.4-2 1.4 (2015). DOI: 10 . 1145 / 2016. DOI: 10.1109/CVPR.2016.91.
2939672.2939785.
[88] John J Hopfield. “Neural networks and physical systems
[75] Horace B Barlow. “Unsupervised learning”. In: Neural with emergent collective computational abilities.” In: Pro-
computation 1.3 (1989), pp. 295–311. DOI: 10 . 1162 / ceedings of the national academy of sciences 79.8 (1982),
neco.1989.1.3.295. pp. 2554–2558. DOI: 10.1073/pnas.79.8.2554.
[76] J MacQueen. “Some Methods for Classification and Anal- [89] Jeffrey L Elman. “Finding structure in time”. In: Cogni-
ysis of Multivariate Observations”. In: Proceedings of the tive science 14.2 (1990), pp. 179–211. DOI: 10 . 1207 /
5th Berkeley Symposium on Mathematical Statistics and s15516709cog1402_1.
Probability. Ed. by Lucien M. Le Cam and Jerzy Ney-
man. University of California Press, 1967. [90] Sepp Hochreiter. “The vanishing gradient problem dur-
ing learning recurrent neural nets and problem solu-
[77] Martin Ester et al. “A Density-Based Algorithm for Dis- tions”. In: International Journal of Uncertainty, Fuzziness
covering Clusters in Large Spatial Databases”. In: Pro- and Knowledge-Based Systems 6.02 (1998), pp. 107–116.
ceedings of the International Conference on Knowledge DOI : 10.1142/S0218488598000094.
Discovery and Data Mining (KDD). Ed. by Hans-Peter
Kriegel and Jörg Sander. AAAI Press, 1996, pp. 226–231. [91] Yoshua Bengio, Patrice Simard, and Paolo Frasconi.
“Learning long-term dependencies with gradient descent
[78] David E Rumelhart, Geoffrey E Hinton, and Ronald J is difficult”. In: IEEE transactions on neural networks 5.2
Williams. “Learning representations by back-propagating (1994), pp. 157–166. DOI: 10.1109/72.279181.
errors”. In: nature 323.6088 (1986), pp. 533–536. DOI:
10.1038/323533a0. [92] Jeremy Appleyard, Tomas Kocisky, and Phil Blunsom.
“Optimizing performance of recurrent neural networks
[79] Frank Rosenblatt. “The perceptron: a probabilistic model on gpus”. In: arXiv preprint (2016). DOI: 10 . 48550 /
for information storage and organization in the brain.” In: arXiv.1604.01946. eprint: 1604.01946. URL: https:
Psychological review 65.6 (1958), p. 386. DOI: 10.1037/ //arxiv.org/abs/1604.01946.
h0042519.
[93] Sepp Hochreiter and Jürgen Schmidhuber. “Long short-
[80] Marius-Constantin Popescu et al. “Multilayer perceptron term memory”. In: Neural computation 9.8 (1997),
and neural networks”. In: WSEAS Transactions on Cir- pp. 1735–1780.
cuits and Systems 8.7 (2009), pp. 579–588.
[94] Geoffrey E Hinton and Ruslan R Salakhutdinov. “Reduc-
[81] Wei Zhang et al. “Shift-invariant pattern recognition neu- ing the dimensionality of data with neural networks”. In:
ral network and its optical architecture”. In: Proceedings science 313.5786 (2006), pp. 504–507. DOI: 10.1126/
of the Annual Conference of the Japan Society of Ap- science.1127647.
plied Physics. Ed. by Japan Society of Applied Physics.
Vol. 564. Japan Society of Applied Physics. Tokyo, Japan, [95] Chuanmin Jia et al. “Layered image compression us-
1988. ing scalable auto-encoder”. In: 2019 IEEE Conference
on Multimedia Information Processing and Retrieval
[82] Yann LeCun et al. “Backpropagation applied to hand- (MIPR). Ed. by IEEE Conference on Multimedia Infor-
written zip code recognition”. In: Neural computation 1.4 mation Processing and Retrieval (MIPR). IEEE. 2019,
(1989), pp. 541–551. DOI: 10.1162/neco.1989.1.4. pp. 431–436. DOI: 10.1109/CVPR.2016.91.
541.
[96] Thorben Finke et al. “Autoencoders for unsupervised
[83] Kevin J Lang, Alex H Waibel, and Geoffrey E Hinton. “A anomaly detection in high energy physics”. In: Journal
time-delay neural network architecture for isolated word of High Energy Physics 2021.6 (2021), pp. 1–32. DOI:
recognition”. In: Neural networks 3.1 (1990), pp. 23–43. 10.1007/JHEP06(2021)161.
DOI : 10.1016/0893-6080(90)90044-L.
[97] Andri Ashfahani et al. “DEVDAN: Deep evolving de-
[84] Karen Simonyan and Andrew Zisserman. “Very deep con- noising autoencoder”. In: Neurocomputing 390 (2020),
volutional networks for large-scale image recognition”. pp. 297–314. DOI: 10.1016/j.neucom.2019.07.106.
In: arXiv preprint (2014). DOI: 10.48550/arXiv.1409.
1556. arXiv: 1409.1556 [cs.CV]. [98] Ian Goodfellow et al. “Generative adversarial nets”. In:
Advances in neural information processing systems 27
(2014).

37
 [99] Charles Miller Grinstead and James Laurie Snell. In- [114] Diana Borsa, Thore Graepel, and John Shawe-Taylor.
troduction to probability. American Mathematical Soc., “Learning shared representations in multi-task reinforce-
1997. ISBN: 188652923X. ment learning”. In: arXiv preprint arXiv:1603.02041
[100] Junyoung Oh, Kyungha Min, and Heekyung Yang. “A (2016). DOI: 10.48550/arXiv.1603.02041.
Pose and Style-Invariant Reenactment Technique for [115] Sinno Jialin Pan and Qiang Yang. “A survey on trans-
Artistic Portraits Using GAN”. In: IEEE Access 10 fer learning”. In: IEEE Transactions on knowledge and
(2022), pp. 50351–50362. DOI: 10 . 1109 / ACCESS . data engineering 22.10 (2009), pp. 1345–1359. DOI: 10.
2022.3172297. 1109/TKDE.2009.191.
[101] Christian Ledig et al. “Photo-realistic single image super- [116] Liviu Panait and Sean Luke. “Cooperative multi-agent
resolution using a generative adversarial network”. In: learning: The state of the art”. In: Autonomous agents and
Proceedings of the International Conference on Computer multi-agent systems 11 (2005), pp. 387–434. DOI: 10 .
Vision. Ed. by ICCV. Springer, 2017, pp. 90–99. 1007/s10458-005-2631-2.
[102] Xintao Wang et al. “Esrgan: Enhanced super-resolution [117] Karl Weiss, Taghi M Khoshgoftaar, and DingDing Wang.
generative adversarial networks”. In: Proceedings of the “A survey of transfer learning”. In: Journal of Big data 3
European conference on computer vision (ECCV) 2018 (2016), pp. 1–40. DOI: 10.1186/S40537-016-0043-6.
workshops. Ed. by Springer. 2018, pp. 0–0. DOI: 10 .
[118] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben-
1007/978-3-030-11021-5_5.
gio. “Neural machine translation by jointly learning
[103] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. to align and translate”. In: arXiv preprint (2014).
“Generating videos with scene dynamics”. In: Advances arXiv:1409.0473. DOI: 10.48550/arXiv.1409.0473.
in neural information processing systems 29 (2016).
[119] Ashish Vaswani et al. “Attention is all you need”. In:
[104] Aayush Bansal et al. “Recycle-GAN: Unsupervised Video Advances in neural information processing systems 30
Retargeting”. In: Proceedings of the European Confer- (2017).
ence on Computer Vision (ECCV). Ed. by Vittorio Fer-
[120] Alexey Dosovitskiy et al. “An image is worth 16x16
rari et al. Vol. 11205. Lecture Notes in Computer Science.
words: Transformers for image recognition at scale”. In:
Springer. Munich, Germany, 2018, pp. 119–135. DOI: 10.
arXiv preprint (2020). DOI: 10 . 48550 / arXiv . 2010 .
1007/978-3-030-01246-5_8.
11929. URL: https://arxiv.org/abs/2010.11929.
[105] Richard S Sutton. “Learning to predict by the methods
[121] Colin Raffel et al. “Exploring the limits of transfer learn-
of temporal differences”. In: Machine learning 3 (1988),
ing with a unified text-to-text transformer”. In: Journal of
pp. 9–44. DOI: 10.1007/BF00115009.
machine learning research 21.140 (2020), pp. 1–67.
[106] Alex M Andrew. “Reinforcement learning:: An introduc-
[122] Alec Radford et al. “Improving language understanding
tion”. In: Kybernetes 27.9 (1998), pp. 1093–1096.
by generative pre-training”. In: xxx (2018).
[107] Christopher JCH Watkins and Peter Dayan. “Q-learning”.
In: Machine learning 8 (1992), pp. 279–292. DOI: 10 . [123] Timothy O Hodson. “Root mean square error (RMSE) or
1007/BF00992698. mean absolute error (MAE): When to use them or not”.
In: Geoscientific Model Development Discussions 2022
[108] John Schulman et al. “Proximal policy optimization (2022), pp. 1–10. DOI: 10.5194/gmd-15-5481-2022.
algorithms”. In: arXiv (2017). Preprint available at
arXiv:1707.06347. DOI: 10 . 48550 / arXiv . 1707 . [124] Vagelis Plevris et al. “Investigation of performance met-
06347. arXiv: 1707.06347 [cs.LG]. rics in regression analysis and machine learning-based
prediction models”. In: Proceedings of the 8th European
[109] H. Brendan McMahan and Eider Moore. Congress on Computational Methods in Applied Sciences
“Communication-Efficient Learning of Deep Networks and Engineering (ECCOMAS Congress 2022). Ed. by
from Decentralized Data”. In: Proceedings of the 34th Trond Kvamsdal et al. European Community on Com-
International Conference on Machine Learning (ICML). putational Methods in Applied Sciences. Oslo, Norway,
Ed. by Doina Precup and Yee Whye Teh. PMLR, 2017, 2022. DOI: 10.23967/eccomas.2022.155.
pp. 254–262.
[125] Muhammad Khurram Jassal. The Effect of Optimization
[110] Peter Kairouz et al. “Advances and open problems in of Error Metrics. Accessed: October 03, 2024. 2011.
federated learning”. In: Foundations and trends in ma-
chine learning 14.1–2 (2021), pp. 1–210. DOI: 10.1561/ [126] Sewall Wright. “Correlation and causation”. In: Journal
2200000083. of agricultural research 20.7 (1921), p. 557.
[111] Jiquan Ngiam and Aditya Khosla. “Multimodal Learn- [127] James T Townsend. “Theoretical analysis of an alphabetic
ing Applications in Satellite Imagery”. In: Proceedings of confusion matrix”. In: Perception and Psychophysics 9
the IEEE International Conference on Computer Vision (1971), pp. 40–50. DOI: 10.3758/BF03213026.
(ICCV). Ed. by ICCV. IEEE, 2011, pp. 495–500. [128] Marina Sokolova and Guy Lapalme. “A systematic anal-
[112] Navonil Majumder et al. “Multimodal sentiment analy- ysis of performance measures for classification tasks”.
sis using hierarchical fusion with context modeling”. In: In: Information processing and management 45.4 (2009),
Knowledge-based systems 161 (2018), pp. 124–133. DOI: pp. 427–437. DOI: 10.1016/j.ipm.2009.03.002.
10.1016/j.knosys.2018.07.041. [129] Paul Jaccard. “Etude comparative de la distribution florale
[113] Said Yacine Boulahia et al. “Early, intermediate and late dans une portion des Alpes et des Jura”. In: Bull Soc Vau-
fusion strategies for robust deep learning-based multi- doise Sci Nat 37 (1901), pp. 547–579. DOI: 10 . 5169 /
modal action recognition”. In: Machine Vision and Ap- seals-266450.
plications 32.6 (2021), p. 121. DOI: 10.1007/s00138-
021-01249-8.
38
[130] Beinan Wang. “A Parallel Implementation of Comput- [145] Darius Lam et al. “xView: Objects in Context in Overhead
ing Mean Average Precision”. In: arXiv (2022). DOI: Imagery”. In: arXiv (2018). DOI: 10 . 48550 / arXiv .
10 . 48550 / arXiv . 2206 . 09504. eprint: 2206 . 09504 1802.07856. arXiv: 1802.07856 [cs.CV].
(cs.CV). [146] Gupta Ritwik et al. “xbd: A dataset for assessing build-
[131] Matej Kristan et al. “The eighth visual object tracking ing damage from satellite imagery”. In: arXiv preprint
VOT2020 challenge results”. In: Computer Vision–ECCV (2019), pp. 1–9. DOI: 10.48550/arXiv.1911.09296.
2020 Workshops: Glasgow, UK, August 23–28, 2020, Pro- [147] Fernando Paolo et al. “xview3-sar: Detecting dark fish-
ceedings, Part V 16. Ed. by ECCV 2020. Springer. 2020, ing activity using synthetic aperture radar imagery”. In:
pp. 547–601. DOI: 10.1007/978-3-030-68238-5_39. Advances in Neural Information Processing Systems 35
[132] John A Richards, John A Richards, et al. Remote sens- (2022), pp. 37604–37616. DOI: 10 . 48550 / arXiv .
ing digital image analysis. Vol. 5. Springer, 2022. DOI: 2206.00897.
10.1007/978-3-030-82327-6. [148] Assal Farrahi and Isabel Grande. “Simplified analy-
[133] P. Shanmugapriya et al. “Applications of remote sensing sis of the thermal behavior of a spinning satellite fly-
in agriculture - A review”. In: International Journal of ing over Sun-synchronous orbits”. In: Applied Thermal
Current Microbiology and Applied Sciences 8.01 (2019), Engineering 125 (July 2017). DOI: 10 . 1016 / j .
pp. 2270–2283. DOI: 10 . 20546 / ijcmas . 2019 . 801 . applthermaleng.2017.07.033.
238. [149] Gui-Song Xia et al. “AID: A benchmark data set for per-
[134] David L Verbyla. Satellite remote sensing of natural re- formance evaluation of aerial scene classification”. In:
sources. CRC Press, 2022. ISBN: 9780138740191. DOI: IEEE Transactions on Geoscience and Remote Sensing
10.1201/9780138740191. 55.7 (2017), pp. 3965–3981. DOI: 10.1109/TGRS.2017.
[135] Maik Netzband, William L Stefanov, and Charles Red- 2685945.
man. Applied remote sensing for urban planning, gov- [150] UCAR Climate Data Guide. National Agriculture Im-
ernance and sustainability. Springer Science & Business agery Program (NAIP) Hub Site. Accessed: October 03,
Media, 2007. DOI: 10.1007/978-3-540-68009-3. 2024. 2024. URL: https://naip- usdaonline.hub.
[136] NASA EOSDIS. LP DAAC: Land Processes Distributed arcgis.com/.
Active Archive Center. Accessed: October 03, 2024. 2023. [151] Yi Yang and Shawn Newsam. “Bag-of-visual-words and
URL: https://lpdaac.usgs.gov/. spatial extensions for land-use classification”. In: Pro-
[137] NASA EOSDIS. GHRC DAAC: Global Hydrology Re- ceedings of the 18th SIGSPATIAL International Confer-
source Center Distributed Active Archive Center. Ac- ence on Advances in Geographic Information Systems.
cessed: October 03, 2024. 2023. URL: https://ghrc. Ed. by SIGSPATIAL. 2010, pp. 270–279. DOI: 10.1145/
nsstc.nasa.gov/. 1869790.1869829.
[138] Jet Propulsion Laboratory (JPL). NSIDC DAAC - Na- [152] Ye Lyu et al. “UAVid: A semantic segmentation dataset
tional Snow and Ice Data Center. Accessed: October 03, for UAV imagery”. In: ISPRS journal of photogramme-
2024. 2023. URL: https://nsidc.org/daac. try and remote sensing 165 (2020), pp. 108–119. DOI:
10.1016/j.isprsjprs.2020.05.009.
[139] Eli Shusterman. NASA Goddard Earth Sciences Data and
Information Services Center (GES DISC). Accessed: Oc- [153] Jet Propulsion Laboratory (JPL). UAVSAR - Uninhabited
tober 03, 2024. 2024. URL: https://disc.gsfc.nasa. Aerial Vehicle Synthetic Aperture Radar. Accessed: Octo-
gov/. ber 03, 2024. 2023. URL: https://uavsar.jpl.nasa.
gov/.
[140] NASA PO.DAAC. PO.DAAC (Physical Oceanography
Distributed Active Archive Center). Accessed: October [154] European Space Agency. Copernicus DEM. Accessed:
03, 2024. 2024. URL: https : / / podaac . jpl . nasa . October 03, 2024. 2022. DOI: 10.5270/esa- c5d3d65.
gov/. URL : http://dx.doi.org/10.5270/ESA-c5d3d65.

[141] Copernicus Programme. Past, Present, and Future of the [155] U.S. Geological Survey. USGS EarthExplorer. Ac-
Satellite-Based Automatic Identification System: Areas of cessed: October 03, 2024. 2024. URL: https : / /
Applications (2004–2016). Accessed: October 03, 2024. earthexplorer.usgs.gov/.
2024. URL: https : / / www . copernicus . eu / en / [156] OpenTopography contributors. OpenTopography: High-
about-copernicus. Resolution Topography Data and Tools. Accessed: Octo-
[142] Jan Musial et al. Overview of the Copernicus Data Space ber 03, 2024. 2024. URL: https://opentopography.
Ecosystem APIs. Tech. rep. Copernicus Meetings, 2024. org/.
DOI : 10.5194/egusphere-egu24-10109. [157] NASA LP DAAC. NASA Global Ecosystem Dynamics In-
[143] Jia Deng et al. “Imagenet: A large-scale hierarchical im- vestigation (GEDI). Accessed: October 03, 2024. 2023.
age database”. In: 2009 IEEE Conference on Computer URL : https : / / lpdaac . usgs . gov / products /
Vision and Pattern Recognition. Ed. by IEEE CVPR. gedi01_bv001/.
IEEE. 2009, pp. 248–255. DOI: 10.1109/CVPR.2009. [158] Hans Karl Heidemann. Lidar base specification. Tech.
5206848. rep. US Geological Survey, 2012.
[144] Wenwen Li et al. “GeoImageNet: a multi-source natural [159] J Hildebrand. UK Environment Agency Data Download.
feature benchmark dataset for GeoAI and supervised ma- Accessed: October 03, 2024. 2023. URL: https : / /
chine learning”. In: GeoInformatica 27.3 (2023), pp. 619– environment.data.gov.uk/.
640. DOI: 10.1007/s10707-022-00476-z.

39
[160] Weikai Tan et al. “Toronto-3D: A Large-Scale Mobile Li- [176] USDA Farm Service Agency. UC Merced Land Use
DAR Dataset for Semantic Segmentation of Urban Road- Dataset on Kaggle. Accessed: October 03, 2024. 2023.
ways”. In: Proceedings of the IEEE/CVF Conference on URL : https : / / www . kaggle . com / datasets /
Computer Vision and Pattern Recognition (CVPR) Work- abdulhasibuddin/uc-merced-land-use-dataset.
shops. Ed. by CVPR Workshop. June 2020. DOI: 10 . [177] Ye Lyu. UAVid: A Semantic Segmentation Dataset for
1109/CVPRW50498.2020.00109. Urban Scenes from UAVs. Accessed: October 03, 2024.
[161] Federal Geographic Data Committee. Marine LiDAR 2024. URL: https://uavid.nl/.
dataset. Accessed: October 03, 2024. 2018. URL: https: [178] Jet Propulsion Laboratory (JPL). Jet Propulsion Labo-
//datasets.seed.nsw.gov.au/dataset/marine- ratory (JPL). Accessed: October 03, 2024. 2024. URL:
lidar-topo-bathy-2018. https://www.jpl.nasa.gov.
[162] OpenStreetMap contributors. OpenStreetMap official [179] Peter L Guth et al. “Digital elevation models: Terminol-
website. Accessed: October 03, 2024. 2024. URL: https: ogy and definitions”. In: Remote Sensing 13.18 (2021),
//www.openstreetmap.org. p. 3581. DOI: 10.3390/rs13183581.
[163] Peter Mooney. Google Maps. Accessed: October 03, [180] Scott Rayburg, Martin Thoms, and Melissa Neave. “A
2024. 2024. URL: https://www.google.com/maps. comparison of digital elevation models generated from
[164] Microsoft Research. Microsoft Geolife GPS Trajectory different data sources”. In: Geomorphology 106.3-4
Dataset. Accessed: October 03, 2024. 2012. URL: https: (2009), pp. 261–270. DOI: 10 . 1016 / j . geomorph .
/ / www . kaggle . com / datasets / arashnic / 2008.11.007.
microsoft-geolife-gps-trajectory-dataset. [181] Natural Resources Canada. Elevation data. Accessed:
[165] MarineTraffic and ExactEarth. MarineTraffic and Ex- October 03, 2024. 2024. URL: https : / / natural -
actEarth AIS Dataset. Accessed: October 03, 2024. resources . canada . ca / maps - tools - and -
2024. URL: https : / / www . marinetraffic . org / publications / satellite - imagery - elevation -
exactearth. data-and-air-photos/elevation-data/24975.
[166] Michael Burri et al. “The EuRoC micro aerial vehicle [182] Alex Shustrik. Elevation Modeling – which to choose:
datasets”. In: The International Journal of Robotics Re- DTM, DSM or DEM? Accessed: October 03, 2024. 2024.
search (2016). DOI: 10 . 1177 / 0278364915620033. URL : https : / / shustrik - maps . com / elevation -
eprint: http://ijr.sagepub.com/content/early/ modeling-which-to-choose-dtm-dsm-or-dem/.
2016/01/21/0278364915620033.full.pdf+html. [183] Federal Geographic Data Committee et al. “National stan-
URL: http://ijr.sagepub.com/content/early/ dard for spatial data accuracy (NSSDA)”. In: Federal Ge-
2016/01/21/0278364915620033. ographic Data Committee, Washington, DC (1998).
[167] Peter Mooney. GeoNames Geographical Database. Ac- [184] Michael Abrams et al. “The aster global dem”. In:
cessed: October 03, 2024. 2024. URL: https : / / www . Photogrammetric Engineering and Remote Sensing 76.4
geonames.org/. (2010), pp. 344–348.
[168] Peter Mooney. Australian Government Data Portal. Ac- [185] NASA Earth Science Data and Information System (ES-
cessed: October 03, 2024. 2024. URL: https://data. DIS). NASA Earthdata. Accessed: October 03, 2024.
gov.au/. 2024. URL: https://earthdata.nasa.gov/.
[169] Institut national de l’information géographique et [186] Ghjulia Sialelli et al. “AGBD: A Global-scale Biomass
forestière (IGN). adresse.data.gouv.fr - Base Adresse Na- Dataset”. In: arXiv preprint (2024). DOI: 10 . 48550 /
tionale (BAN). Accessed: October 03, 2024. 2024. URL: arXiv.2406.04928. arXiv: 2406.04928 [cs.DS].
https://adresse.data.gouv.fr/.
[187] RO Dubayah et al. GEDI L4A Footprint Level Above-
[170] Peter Mooney. ISPARK Parking Dataset. Accessed: Oc- ground Biomass Density, Version 2.1. ORNL DAAC, Oak
tober 03, 2024. 2024. URL: https://data.ibb.gov. Ridge, Tennessee, USA. 2022.
tr/en/dataset/ispark-otopark-bilgileri.
[188] Daniel Callahan and Mike Mustafa Berber. “Vertical ac-
[171] LIN Zongjian. “UAV for mapping—low altitude pho- curacy of the USGS 3DEP program data: study cases
togrammetric survey”. In: International Archives of Pho- in Fresno County and in Davis, California”. In: Boletim
togrammetry and Remote Sensing, Beijing, China 37 de Ciencias Geodesicas 28 (2022), e2022004. DOI: 10.
(2008), pp. 1183–1186. 1590/s1982-21702022000100004.
[172] Wuhan University. AID: A Benchmark Dataset for Per- [189] Microsoft. GlobalMLBuildingFootprints: Worldwide
formance Evaluation of Aerial Scene Classification. Ac- Building Footprints Derived from Satellite Imagery. Ac-
cessed: October 03, 2024. 2023. URL: https : / / cessed: October 03, 2024. 2024. URL: https://github.
captain-whu.github.io/AID/. com/microsoft/GlobalMLBuildingFootprints.
[173] USDA Farm Service Agency. Google Earth Engine. Ac- [190] Maxar Technologies. Prediction of Maize Crop Coeffi-
cessed: October 03, 2024. 2023. URL: https : / / www . cient from UAV Multisensor Remote Sensing Using Ma-
fsa.usda.gov. chine Learning Methods. Accessed: October 03, 2024.
[174] Assal Farrahi. Earthdata - USGS. Accessed: October 03, 2024. URL: https://www.maxar.com.
2024. 2024. URL: https://earthdata.usgs.gov/. [191] Airbus Defence and Space. Prediction of maize crop coef-
[175] Assal Farrahi. Google Earth Engine. Accessed: October ficient from UAV multisensor remote sensing using ma-
03, 2024. 2024. URL: https://earthengine.google. chine learning methods. Accessed: October 03, 2024.
com/. 2023. URL: https : / / www . airbus . com / en /
products-services/defence-and-space.
40
[192] Institut national de l’information géographique et [209] Norman Einspruch. Application specific integrated circuit
forestière (IGN). Prediction of maize crop coefficient from (ASIC) technology. Vol. 23. Academic Press, 2012. ISBN:
UAV multisensor remote sensing using machine learn- 0124315216.
ing methods. Accessed: October 03, 2024. 2024. URL: [210] Sang-Won Lee et al. “A case for flash memory SSD in
https://www.ign.fr. enterprise database applications”. In: Proceedings of the
[193] Peter Mooney, Marco Minghini, et al. “A review of Open- 2008 ACM SIGMOD International Conference on Man-
StreetMap data”. In: Mapping and the citizen sensor agement of Data. Ed. by SIGMOD. 2008, pp. 1075–1086.
(2017), pp. 37–59. DOI: 10.5334/bbf.c. DOI : 10.1145/1376616.1376723.
[194] NASA Earth Science Data. Geofabrik Download Server. [211] Manar Abu Talib et al. “A systematic literature review on
Accessed: October 03, 2024. 2024. URL: https : / / hardware implementation of artificial intelligence algo-
download.geofabrik.de. rithms”. In: The Journal of Supercomputing 77.2 (2021),
[195] NASA Earth Science Data. BBBike Extracts. Accessed: pp. 1897–1938. DOI: 10.1007/s11227-020-03325-8.
October 03, 2024. 2024. URL: https : / / extract . [212] William J. Dally et al. “Hardware-Enabled Artificial Intel-
bbbike.org. ligence”. In: 2018 IEEE Symposium on VLSI Circuits. Ed.
[196] Peter Mooney. Google Earth. Accessed: October 03, by IEEE. IEEE. 2018, pp. 3–6. DOI: 10.1109/VLSIC.
2024. 2024. URL: https://earth.google.com/. 2018.8502368.
[197] Peter Mooney. Google Places API. Accessed: October 03, [213] Mark Zhao et al. “Tectonic-Shift: A Composite Storage
2024. 2024. URL: https://developers.google.com/ Fabric for Large-Scale ML Training”. In: 2023 USENIX
maps/documentation/places. Annual Technical Conference (USENIX ATC 23). Ed. by
USENIX Association. 2023, pp. 433–449. ISBN: 978-1-
[198] Peter Mooney. Google Directions API. Accessed: Oc- 939133-35-9.
tober 03, 2024. 2024. URL: https : / / developers .
google.com/maps/documentation/directions. [214] Satadru Pan et al. “Facebook’s tectonic filesystem: Ef-
ficiency from exascale”. In: 19th USENIX Conference
[199] Microsoft Research. Deep Learning Models Performance on File and Storage Technologies (FAST 21). Ed. by
Evaluations for Remote Sensed Image Classification. Ac- USENIX Association. 2021, pp. 217–231. ISBN: 978-1-
cessed: October 03, 2024. 2023. URL: https : / / www . 939133-17-5.
microsoft.com/en-us/research/.
[215] Ana Klimovic et al. “Flash storage disaggregation”. In:
[200] Mélanie Fournier et al. “Past, present, and future of the Proceedings of the Eleventh European Conference on
satellite-based automatic identification system: Areas of Computer Systems. Ed. by EuroSys. 2016, pp. 1–15. DOI:
applications (2004–2016)”. In: WMU journal of maritime 10.1145/2901318.2901337.
affairs 17.3 (2018), pp. 311–345. DOI: 10 . 1121 / 1 .
4763548. [216] Francis J. Pierce and Peter Nowak. “Aspects of Precision
Agriculture”. In: Advances in Agronomy. Ed. by Donald
[201] Peter Mooney. GeoNames Feature Density Map. Ac- L. Sparks. Vol. 67. Elsevier, 1999, pp. 1–85. ISBN: 978-0-
cessed: October 03, 2024. 2006. URL: https : / / 12-000767-7. DOI: 10.1016/S0065-2113(08)60513-
geonames.wordpress.com/2006/12/07/geonames- 1. (Visited on 03/03/2024).
feature-density-map/.
[217] Jinbang Peng et al. “Combination of UAV and deep learn-
[202] Peter Mooney. Direction Interministérielle du Numérique
ing to estimate wheat yield at ripening stage: The potential
(DINUM) - Site officiel. 2024. URL: https : / / www .
of phenotypic features”. In: International Journal of Ap-
numerique.gouv.fr.
plied Earth Observation and Geoinformation 124 (Nov.
[203] Peter Mooney. Agence Nationale de la Cohésion des Ter- 2023), p. 103494. ISSN: 15698432. DOI: 10 . 1016 / j .
ritoires (ANCT) - Site officiel. 2024. URL: https://www. jag.2023.103494. (Visited on 12/21/2023).
anct.gouv.fr.
[218] Yanyu Wang et al. “Estimation of Rice Growth Param-
[204] Institut national de l’information géographique et eters Based on Linear Mixed-Effect Model Using Multi-
forestière (IGN). Institut national de l’information spectral Images from Fixed-Wing Unmanned Aerial Vehi-
géographique et forestière (IGN) - Site officiel. Accessed: cles”. In: Remote Sensing 11.11 (June 8, 2019), p. 1371.
October 03, 2024. 2024. URL: https://www.ign.fr. ISSN : 2072-4292. DOI : 10.3390/rs11111371. (Visited
[205] John D Owens et al. “GPU computing”. In: Proceedings on 12/22/2023).
of the IEEE 96.5 (2008), pp. 879–899. DOI: 10 . 1109 / [219] Franciele Morlin Carneiro et al. “Soil and satellite remote
JPROC.2008.917757. sensing variables importance using machine learning to
[206] Jonathan Rose, Abbas El Gamal, and Alberto predict cotton yield”. In: Smart Agricultural Technology 5
Sangiovanni-Vincentelli. “Architecture of field- (Oct. 2023), p. 100292. ISSN: 27723755. DOI: 10.1016/
programmable gate arrays”. In: Proceedings of the IEEE j.atech.2023.100292. (Visited on 12/21/2023).
81.7 (1993), pp. 1013–1029. DOI: 10.1109/5.231340. [220] Zeeshan Ramzan et al. “A Multimodal Data Fusion and
[207] Jennifer Eyre and Jeff Bier. “The evolution of DSP Deep Neural Networks Based Technique for Tea Yield Es-
processors”. In: IEEE Signal Processing Magazine 17.2 timation in Pakistan Using Satellite Imagery”. In: IEEE
(2000), pp. 43–51. Access 11 (2023), pp. 42578–42594. ISSN: 2169-3536.
DOI : 10 . 1109 / ACCESS . 2023 . 3271410. (Visited on
[208] Norman P Jouppi et al. “In-datacenter performance anal-
12/21/2023).
ysis of a tensor processing unit”. In: Proceedings of the
44th annual international symposium on computer archi-
tecture. Ed. by ISCA. 2017, pp. 1–12. DOI: 10 . 1145 /
3079856.308024.
41
[221] Anesmar Olino De Albuquerque et al. “Dealing With [233] Hengshuang Zhao et al. “Pyramid Scene Parsing Net-
Clouds and Seasonal Changes for Center Pivot Irriga- work”. In: Proceedings of the IEEE Conference on Com-
tion Systems Detection Using Instance Segmentation in puter Vision and Pattern Recognition (CVPR). Ed. by
Sentinel-2 Time Series”. In: IEEE Journal of Selected IEEE CVPR. 2017, pp. 2881–2890. DOI: 10 . 1109 /
Topics in Applied Earth Observations and Remote Sens- CVPR.2017.660.
ing 14 (2021), pp. 8447–8457. ISSN: 1939-1404, 2151- [234] Lee R Dice. “Measures of the amount of ecologic associ-
1535. DOI: 10.1109/JSTARS.2021.3104726. (Visited ation between species”. In: Ecology 26.3 (1945), pp. 297–
on 12/21/2023). 302. DOI: 10.2307/1932409.
[222] Ehsan Raei et al. “A deep learning image segmentation [235] Toby N Carlson and David A Ripley. “On the relation be-
model for agricultural irrigation system classification”. tween NDVI, fractional vegetation cover, and leaf area
In: Computers and Electronics in Agriculture 198 (July index”. In: Remote sensing of Environment 62.3 (1997),
2022), p. 106977. ISSN: 01681699. DOI: 10 . 1016 / j . pp. 241–252. DOI: 10.1016/S0034-4257(97)00104-
compag.2022.106977. (Visited on 12/21/2023). 1.
[223] Zhu Han et al. “Spatio-temporal multi-level attention crop [236] Harris Drucker et al. “Support vector regression ma-
mapping method using time-series SAR imagery”. In: chines”. In: Advances in neural information processing
ISPRS Journal of Photogrammetry and Remote Sensing systems 9 (1996).
206 (Dec. 2023), pp. 293–310. ISSN: 09242716. DOI:
[237] Ella Bingham and Heikki Mannila. “Random Projection
10 . 1016 / j . isprsjprs . 2023 . 11 . 016. (Visited on
in Dimensionality Reduction: Applications to Image and
12/21/2023).
Text Data”. In: Proceedings of the Seventh ACM SIGKDD
[224] Saskia Wolff et al. Identifying agricultural landscape International Conference on Knowledge Discovery and
types for brandenburg, Germany using IACS data. Tech. Data Mining (KDD). Ed. by Chengqi Zhang, Hans W.
rep. FORLand-working paper, 2020. DOI: 10 . 18452 / Guesgen, and Wai-Kiang Yeap. ACM, 2001, pp. 245–250.
21342. DOI : 10.1145/502512.50254.
[225] VSF Garnot and L Landrieu. Pastis-r-panopiic segmenta- [238] Kaiming He et al. “Mask R-CNN”. In: Proceedings of
tion of radar and optical satellite image time series. Ac- the IEEE International Conference on Computer Vision
cessed: October 03, 2024. 2021. (ICCV). Ed. by IEEE. 2017, pp. 2961–2969. DOI: 10 .
[226] Western Cape Department of Agriculture. Crop type clas- 1109/ICCV.2017.322.
sification dataset for Western Cape, South Africa. Version [239] Cynthia C Bennington and William V Thayne. “Use and
1.0. Accessed: October 03, 2024. 2021. URL: https:// misuse of mixed model analysis of variance in ecologi-
radiantearth.blob.core.windows.net/. cal studies”. In: Ecology 75.3 (1994), pp. 717–722. DOI:
[227] Tsung-Yi Lin et al. “Feature Pyramid Networks for Ob- 10.2307/1941729.
ject Detection”. In: Proceedings of the IEEE Conference [240] Guang Zheng and L Monika Moskal. “Retrieving leaf
on Computer Vision and Pattern Recognition (CVPR). Ed. area index (LAI) using remote sensing: theories, methods
by IEEE CVPR. 2017, pp. 2117–2125. DOI: 10.1109/ and sensors”. In: Sensors 9.4 (2009), pp. 2719–2745. DOI:
CVPR.2017.106. 10.3390/s90402719.
[228] Varun Tiwari et al. “Automated in-season rice crop map- [241] Emmanuel Abiodun Abioye et al. “A review on monitor-
ping using Sentinel time-series data and Google Earth En- ing and advanced control strategies for precision irriga-
gine: A case study in climate-risk prone Bangladesh”. In: tion”. In: Computers and Electronics in Agriculture 173
Journal of Environmental Management 351 (Feb. 2024), (June 2020), p. 105441. ISSN: 01681699. DOI: 10.1016/
p. 119615. ISSN: 03014797. DOI: 10.1016/j.jenvman. j.compag.2020.105441. (Visited on 03/03/2024).
2023.119615. (Visited on 12/21/2023).
[242] Saining Xie et al. “Aggregated Residual Transformations
[229] Ping-Sung Liao, Tse-Sheng Chen, Pau-Choo Chung, et al. for Deep Neural Networks”. In: Proceedings of the IEEE
“A Fast Algorithm for Multilevel Thresholding”. In: Jour- Conference on Computer Vision and Pattern Recognition
nal of Information Science and Engineering 17.5 (2001), (CVPR). Ed. by IEEE CVPR. 2017, pp. 1492–1500. DOI:
pp. 713–727. DOI: 10.1688/JISE.2001.17.5.1. 10.1109/CVPR.2017.634.
[230] Sina Mohammadi, Mariana Belgiu, and Alfred Stein. [243] Harikrishnan Muraleedharan Jalajamony et al. “Drone
“Improvement in crop mapping from satellite image time Aided Thermal Mapping for Selective Irrigation of Lo-
series by effectively supervising deep neural networks”. calized Dry Spots”. In: IEEE Access 11 (2023), pp. 7320–
In: ISPRS Journal of Photogrammetry and Remote Sens- 7335. ISSN: 2169-3536. DOI: 10.1109/ACCESS.2023.
ing 198 (Apr. 2023), pp. 272–283. ISSN: 09242716. DOI: 3237546. (Visited on 12/21/2023).
10 . 1016 / j . isprsjprs . 2023 . 03 . 007. (Visited on [244] Abebaw Alem and Shailender Kumar. “Deep Learn-
12/21/2023). ing Models Performance Evaluations for Remote Sensed
[231] João Batista Ribeiro et al. “Automated detection of sugar- Image Classification”. In: IEEE Access 10 (2022),
cane crop lines from UAV images using deep learning”. pp. 111784–111793. ISSN: 2169-3536. DOI: 10 . 1109 /
In: Information Processing in Agriculture (Apr. 2023), ACCESS.2022.3215264. (Visited on 12/22/2023).
S2214317323000501. ISSN: 22143173. DOI: 10.1016/ [245] Boan Chen et al. “Multi-modal fusion of satellite and
j.inpa.2023.04.001. (Visited on 12/22/2023). street-view images for urban village classification based
[232] Abhishek Chaurasia and Eugenio Culurciello. “LinkNet: on a dual-branch deep neural network”. In: International
Exploiting Encoder Representations for Efficient Seman- Journal of Applied Earth Observation and Geoinfor-
tic Segmentation”. In: 2017 IEEE Visual Communications mation 109 (May 2022), p. 102794. ISSN: 15698432.
and Image Processing (VCIP). Ed. by IEEE VCIP. IEEE. DOI : 10 . 1016 / j . jag . 2022 . 102794. (Visited on
2017, pp. 1–4. DOI: 10.1109/VCIP.2017.8305148. 12/22/2023).
42
[246] Pengcheng Zhang et al. “3D Urban Buildings Extraction [257] Sean P. Kearney et al. “Maintaining accurate, current, ru-
Based on Airborne LiDAR and Photogrammetric Point ral road network data: An extraction and updating rou-
Cloud Fusion According to U-Net Deep Learning Model tine using RapidEye, participatory GIS and deep learn-
Segmentation”. In: IEEE Access 10 (2022), pp. 20889– ing”. In: International Journal of Applied Earth Observa-
20897. ISSN: 2169-3536. DOI: 10.1109/ACCESS.2022. tion and Geoinformation 87 (May 2020), p. 102031. ISSN:
3152744. (Visited on 12/22/2023). 15698432. DOI: 10.1016/j.jag.2019.102031. (Vis-
[247] Alireza Attarhay Tehrani et al. “Predicting solar radiation ited on 12/16/2023).
in the urban area: A data-driven analysis for sustainable [258] Javier Grandio et al. “Multimodal deep learning for
city planning using artificial neural networking”. In: Sus- point cloud panoptic segmentation of railway environ-
tainable Cities and Society 100 (Jan. 2024), p. 105042. ments”. In: Automation in Construction 150 (June 2023),
ISSN : 22106707. DOI : 10.1016/j.scs.2023.105042. p. 104854. ISSN: 09265805. DOI: 10.1016/j.autcon.
(Visited on 12/22/2023). 2023.104854. (Visited on 12/23/2023).
[248] S. El Hajjar et al. “Long Short-Term Memory and At- [259] Jianhua Guo et al. “Nationwide urban tree canopy map-
tention Models for Simulating Urban Densification”. In: ping and coverage assessment in Brazil from high-
Sustainable Cities and Society 98 (Nov. 2023), p. 104852. resolution remote sensing images using deep learning”.
ISSN : 22106707. DOI : 10.1016/j.scs.2023.104852. In: ISPRS Journal of Photogrammetry and Remote Sens-
(Visited on 12/22/2023). ing 198 (Apr. 2023), pp. 1–15. ISSN: 09242716. DOI:
[249] Shuting Sun et al. “GAN-Based LUCC Prediction via 10 . 1016 / j . isprsjprs . 2023 . 02 . 007. (Visited on
the Combination of Prior City Planning Information and 12/23/2023).
Land-Use Probability”. In: IEEE Journal of Selected Top- [260] Shinjita Das, Qian (Chayn) Sun, and Heng Zhou. “GeoAI
ics in Applied Earth Observations and Remote Sensing 14 to implement an individual tree inventory: Framework
(2021), pp. 10189–10198. ISSN: 1939-1404, 2151-1535. and application of heat mitigation”. In: Urban Forestry
DOI : 10 . 1109 / JSTARS . 2021 . 3106481. (Visited on & Urban Greening 74 (Aug. 2022), p. 127634. ISSN:
12/22/2023). 16188667. DOI: 10.1016/j.ufug.2022.127634. (Vis-
[250] Jinxing Hu et al. “A spatiotemporal intelligent framework ited on 12/23/2023).
and experimental platform for urban digital twins”. In: [261] Angus J. Carnegie et al. “Airborne multispectral imagery
Virtual Reality and Intelligent Hardware 5.3 (June 2023), and deep learning for biosecurity surveillance of inva-
pp. 213–231. ISSN: 20965796. DOI: 10.1016/j.vrih. sive forest pests in urban landscapes”. In: Urban Forestry
2022.08.018. (Visited on 12/22/2023). & Urban Greening 81 (Mar. 2023), p. 127859. ISSN:
[251] H. Canli and S. Toklu. “Deep Learning-Based Mobile Ap- 16188667. DOI: 10.1016/j.ufug.2023.127859. (Vis-
plication Design for Smart Parking”. In: IEEE Access 9 ited on 12/23/2023).
(2021), pp. 61171–61183. ISSN: 2169-3536. DOI: 10 . [262] Mingxing Tan and Quoc Le. “EfficientNet: Rethinking
1109/ACCESS.2021.3074887. (Visited on 12/22/2023). Model Scaling for Convolutional Neural Networks”. In:
[252] Nanqi Ye et al. “Urban Commerce Distribution Analy- Proceedings of the 36th International Conference on Ma-
sis Based on Street View and Deep Learning”. In: IEEE chine Learning (ICML). Ed. by Kamalika Chaudhuri and
Access 7 (2019), pp. 162841–162849. ISSN: 2169-3536. Ruslan Salakhutdinov. PMLR, 2019, pp. 6105–6114.
DOI : 10 . 1109 / ACCESS . 2019 . 2951294. (Visited on [263] Street View Google. Google Street View: Explore the
12/22/2023). World. Accessed: October 28, 2024. 2024. URL: https:
[253] Dancho Panovski and Titus Zaharia. “Long and Short- //www.google.com/streetview.
Term Bus Arrival Time Prediction With Traffic Density [264] Fisher Yu and Vladlen Koltun. “Multi-scale context ag-
Matrix”. In: IEEE Access 8 (2020), pp. 226267–226284. gregation by dilated convolutions”. In: arXiv preprint
ISSN : 2169-3536. DOI : 10 . 1109 / ACCESS . 2020 . 1511.07122 (2015). URL: https://arxiv.org/abs/
3044173. (Visited on 12/16/2023). 1511.07122.
[254] Noureen Zafar et al. “Traffic Prediction in Smart Cities [265] Avrim Blum and Tom Mitchell. “Combining Labeled and
Based on Hybrid Feature Space”. In: IEEE Access 10 Unlabeled Data with Co-Training”. In: Proceedings of the
(2022), pp. 134333–134348. ISSN: 2169-3536. DOI: 10. Eleventh Annual Conference on Computational Learning
1109/ACCESS.2022.3231448. (Visited on 12/22/2023). Theory (COLT). Ed. by Judea Pearl. 1998, pp. 92–100.
[255] Yihang Lu et al. “IODRNN - Incremental output decom- DOI : 10.1145/279943.279962.
position for a valid traffic flow prediction with GNSS [266] Phillip Isola et al. “Image-to-Image Translation with Con-
data”. In: Engineering Applications of Artificial Intelli- ditional Adversarial Networks”. In: Proceedings of the
gence 128 (Feb. 2024), p. 107520. ISSN: 09521976. DOI: IEEE Conference on Computer Vision and Pattern Recog-
10 . 1016 / j . engappai . 2023 . 107520. (Visited on nition (CVPR). Ed. by IEEE CVPR. 2017, pp. 1125–1134.
12/22/2023). DOI : 10.1109/CVPR.2017.632.
[256] Lei Chen et al. “Personalized itinerary recommendation: [267] Jane Bromley et al. “Signature verification using a
Deep and collaborative learning with textual informa- siamese time delay neural network”. In: Advances in neu-
tion”. In: Expert Systems with Applications 144 (Apr. ral information processing systems 6 (1993).
2020), p. 113070. ISSN: 09574174. DOI: 10 . 1016 / j .
eswa.2019.113070. (Visited on 12/16/2023).

43
[268] Donghao Li et al. “MPSiam: A Fast Multiplexing Siamese [280] Chris Brunsdon, Stewart Fotheringham, and Martin
Tracking Network”. In: Advances in Smart Vehicular Charlton. “Geographically weighted regression”. In:
Technology, Transportation, Communication and Appli- Journal of the Royal Statistical Society: Series D (The
cations: Proceedings of VTCA 2021. Ed. by Palaiah- Statistician) 47.3 (1998), pp. 431–443. DOI: 10 . 1111 /
nakote Shivakumara, Cheng-Lin Tan, and M. H. Supriya. 1467-9884.00145.
Springer, 2021, pp. 295–302. DOI: 10.1007/978-981- [281] Xiaoli Wei et al. “Satellite remote sensing of aerosol
16-4039-1_29. optical depth: Advances, challenges, and perspectives”.
[269] Yimo Wang et al. “Multiple Stream Oriented Siamese In: Critical Reviews in Environmental Science and Tech-
Network for RGB-T Tracking”. In: 2021 13th Interna- nology 50.16 (2020), pp. 1640–1725. DOI: 10 . 1080 /
tional Conference on Wireless Communications and Sig- 10643389.2019.1665944.
nal Processing (WCSP). Ed. by IEEE WCSP. IEEE. 2021, [282] Peiqing Lou et al. “Deep learning reveals rapid vegeta-
pp. 1–5. DOI: 10.1109/WCSP52459.2021.9613242. tion greening in changing climate from 1988 to 2018 on
[270] Guanhao Yang et al. “Face Mask Recognition System the Qinghai-Tibet Plateau”. In: Ecological Indicators 148
with YOLOv5 Based on Image Recognition”. In: 2020 (Apr. 2023), p. 110020. ISSN: 1470160X. DOI: 10.1016/
IEEE 6th International Conference on Computer and j.ecolind.2023.110020. (Visited on 12/19/2023).
Communications (ICCC). Ed. by IEEE ICCC. IEEE. [283] Shidong Wang et al. “Applying deep-learning enhanced
2020, pp. 1398–1404. DOI: 10.1109/ICCC51575.2020. fusion methods for improved NDVI reconstruction and
9345042. long-term vegetation cover study: A case of the Danjiang
[271] Shaoqing Ren et al. “Faster R-CNN: Towards real-time River Basin”. In: Ecological Indicators 155 (Nov. 2023),
object detection with region proposal networks”. In: IEEE p. 111088. ISSN: 1470160X. DOI: 10.1016/j.ecolind.
transactions on pattern analysis and machine intelligence 2023.111088. (Visited on 12/19/2023).
39.6 (2016), pp. 1137–1149. DOI: 10 . 1109 / TPAMI .
[284] Qinchuan Xin et al. “Evaluations and comparisons of
2016.2577031.
rule-based and machine-learning-based methods to re-
[272] Pascal Vincent et al. “Stacked denoising autoencoders: trieve satellite-based vegetation phenology using MODIS
Learning useful representations in a deep network with and USA National Phenology Network data”. In: Interna-
a local denoising criterion.” In: Journal of machine learn- tional Journal of Applied Earth Observation and Geoin-
ing research 11.12 (2010). formation 93 (Dec. 2020), p. 102189. ISSN: 15698432.
[273] Marjan Laal and Seyed Mohammad Ghodsi. “Benefits of DOI : 10 . 1016 / j . jag . 2020 . 102189. (Visited on
collaborative learning”. In: Procedia-social and behav- 12/19/2023).
ioral sciences 31 (2012), pp. 486–490. DOI: 10 . 1016 / [285] Ronald Opio et al. “Modeling the atmospheric disper-
j.sbspro.2011.12.091. sion of SO2 from Mount Nyiragongo”. In: Journal of
[274] Erik Rúa et al. “Automatic detection to inventory road African Earth Sciences 197 (Jan. 2023), p. 104771. ISSN:
slopes using open LiDAR point clouds”. In: Interna- 1464343X. DOI: 10 . 1016 / j . jafrearsci . 2022 .
tional Journal of Applied Earth Observation and Geoin- 104771. (Visited on 12/19/2023).
formation 118 (Apr. 2023), p. 103225. ISSN: 15698432. [286] Minsu Kim, Dominik Brunner, and Gerrit Kuhlmann.
DOI : 10 . 1016 / j . jag . 2023 . 103225. (Visited on “Importance of satellite observations for high-resolution
12/22/2023). mapping of near-surface ISPRS2 by machine learning”.
[275] Vijay Badrinarayanan, Alex Kendall, and Roberto In: Remote Sensing of Environment 264 (Oct. 2021),
Cipolla. “Segnet: A deep convolutional encoder-decoder p. 112573. ISSN: 00344257. DOI: 10 . 1016 / j . rse .
architecture for image segmentation”. In: IEEE transac- 2021.112573. (Visited on 12/19/2023).
tions on pattern analysis and machine intelligence 39.12 [287] Xinyu Dou et al. “Estimates of daily ground-level NO2
(2017), pp. 2481–2495. DOI: 10 . 1109 / TPAMI . 2016 . concentrations in China based on Random Forest model
2644615. integrated K-means”. In: Advances in Applied Energy 2
[276] Charles Ruizhongtai Qi et al. “Pointnet++: Deep hierar- (May 2021), p. 100017. ISSN: 26667924. DOI: 10.1016/
chical feature learning on point sets in a metric space”. j.adapen.2021.100017. (Visited on 12/23/2023).
In: Advances in neural information processing systems 30 [288] Andrew Rowley and Oktay Karakuş. “Predicting air qual-
(2017). ity via multimodal AI and satellite imagery”. In: Re-
[277] Liang-Chieh Chen et al. “Encoder-Decoder with Atrous mote Sensing of Environment 293 (Aug. 2023), p. 113609.
Separable Convolution for Semantic Image Segmenta- ISSN : 00344257. DOI : 10.1016/j.rse.2023.113609.
tion”. In: Proceedings of the European Conference on (Visited on 12/19/2023).
Computer Vision (ECCV). Ed. by Vittorio Ferrari et al. [289] Dewen Seng et al. “Spatiotemporal prediction of air qual-
2018, pp. 801–818. DOI: 10.1007/978-3-030-01234- ity based on LSTM neural network”. In: Alexandria Engi-
2_49. neering Journal 60.2 (Apr. 2021), pp. 2021–2032. ISSN:
[278] Christian Szegedy et al. “Inception-v4, Inception-ResNet 11100168. DOI: 10.1016/j.aej.2020.12.009. (Vis-
and the Impact of Residual Connections on Learning”. ited on 12/19/2023).
In: Proceedings of the Thirty-First AAAI Conference on [290] Lorenzo Gianquintieri et al. “Implementation of a GeoAI
Artificial Intelligence (AAAI). Ed. by AAAI. Vol. 31. 1. model to assess the impact of agricultural land on the
AAAI Press, 2017. DOI: 10.1109/ICCC51575.2020. spatial distribution of PM2.5 concentration”. In: Chemo-
9345042. sphere 352 (Mar. 2024), p. 141438. ISSN: 00456535. DOI:
[279] Israel Cohen et al. “Pearson correlation coefficient”. In: 10.1016/j.chemosphere.2024.141438. (Visited on
Noise reduction in speech processing (2009), pp. 1–4. 03/19/2024).
DOI : 10.1007/978-3-642-00296-0_5.
44
[291] Valeriy Shevchenko et al. “Climate Change Impact on [305] A Stewart Fotheringham, Wenbai Yang, and Wei
Agricultural Land Suitability: An Interpretable Machine Kang. “Multiscale geographically weighted regression
Learning-Based Eurasia Case Study”. In: IEEE Access (MGWR)”. In: Annals of the American Association of Ge-
12 (2024), pp. 15748–15763. ISSN: 2169-3536. DOI: 10. ographers 107.6 (2017), pp. 1247–1265. DOI: 10.1080/
1109/ACCESS.2024.3358865. (Visited on 03/20/2024). 24694452.2017.1352480.
[292] Jing Cao. “The Ecological Safety Assessment and Brand [306] Taylor M Oshan et al. “mgwr: A Python implemen-
Communication of Ice-Snow Tourism Under the Inter- tation of multiscale geographically weighted regression
net of Things and Deep Learning”. In: IEEE Access 11 for investigating process spatial heterogeneity and scale”.
(2023), pp. 128235–128244. ISSN: 2169-3536. DOI: 10. In: ISPRS International Journal of Geo-Information 8.6
1109/ACCESS.2023.3332688. (Visited on 03/19/2024). (2019), p. 269. DOI: 10.3390/ijgi8060269.
[293] Isabelle Guyon et al. “Gene selection for cancer clas- [307] Scott M Lundberg and Su-In Lee. “A unified approach to
sification using support vector machines”. In: Machine interpreting model predictions”. In: Advances in neural
learning 46 (2002), pp. 389–422. DOI: 10 . 1023 / A : information processing systems 30 (2017).
1012487302797. [308] Liudmila Prokhorenkova et al. “CatBoost: unbiased
[294] Yulun Zhang et al. “Image Super-Resolution Using Very boosting with categorical features”. In: Advances in neu-
Deep Residual Channel Attention Networks”. In: Pro- ral information processing systems 31 (2018).
ceedings of the European Conference on Computer Vision [309] Guolin Ke et al. “Lightgbm: A highly efficient gradient
(ECCV). Ed. by Vittorio Ferrari et al. 2018, pp. 286–301. boosting decision tree”. In: Advances in neural informa-
DOI : 10.1007/978-3-030-01234-2_18.
tion processing systems 30 (2017).
[295] Julio L Betancourt et al. Implementing a US national phe- [310] Xingjian Shi et al. “Convolutional LSTM network: A ma-
nology network. Accessed: October 03, 2024. 2005. DOI: chine learning approach for precipitation nowcasting”. In:
10.1029/2005EO510005. Advances in neural information processing systems 28
[296] Han Ma and Shunlin Liang. “Development of the GLASS (2015).
250-m leaf area index product (version 6) from MODIS [311] PS Thenkabail et al. “NASA making earth system data
data using the bidirectional LSTM deep learning model”. records for use in research environments (MEaSUREs)
In: Remote Sensing of Environment 273 (May 2022), global food security support analysis data (GFSAD) Crop
p. 112985. ISSN: 00344257. DOI: 10 . 1016 / j . rse . dominance 2010 global 1 km V001”. In: NASA EOSDIS
2022.112985. (Visited on 12/19/2023). Land Processes DAAC (2016). URL: https : / / nasa .
[297] Mike Schuster and Kuldip K Paliwal. “Bidirectional re- gov/earth-system-data.
current neural networks”. In: IEEE transactions on Signal [312] Ronald A Fisher. “On the Mathematical Foundations
Processing 45.11 (1997), pp. 2673–2681. DOI: 10.1109/ of Theoretical Statistics”. In: Philosophical Transactions
78.650093. 222.594-604 (1922), pp. 309–368. DOI: 10.1098/rsta.
[298] Jiyunting Sun et al. “Aerosol Absorption Over Land De- 1922.0009.
rived From the Ultra-Violet Aerosol Index by Deep Learn-
[313] Jacob Wolfowitz. “The minimum distance method”. In:
ing”. In: IEEE Journal of Selected Topics in Applied Earth
The Annals of Mathematical Statistics (1957), pp. 75–88.
Observations and Remote Sensing 14 (2021), pp. 9692–
9710. ISSN: 1939-1404, 2151-1535. DOI: 10 . 1109 / [314] Ling Du et al. “Drainage ditch network extraction from
JSTARS.2021.3108669. (Visited on 12/19/2023). lidar data using deep convolutional neural networks in a
low relief landscape”. In: Journal of Hydrology 628 (Jan.
[299] O Torres et al. “Derivation of aerosol properties from
2024), p. 130591. ISSN: 00221694. DOI: 10 . 1016 / j .
satellite measurements of backscattered ultraviolet radi-
jhydrol.2023.130591. (Visited on 03/15/2024).
ation: Theoretical basis”. In: Journal of Geophysical Re-
search: Atmospheres 103.D14 (1998), pp. 17099–17110. [315] Shuting Sun et al. “Quadtree decomposition-based Deep
DOI : 10.1029/98JD00900. learning method for multiscale coastline extraction with
high-resolution remote sensing imagery”. In: Science
[300] Brent N Holben et al. “AERONET—A federated instru-
of Remote Sensing 9 (June 2024), p. 100112. ISSN:
ment network and data archive for aerosol characteriza-
26660172. DOI: 10.1016/j.srs.2023.100112. (Vis-
tion”. In: Remote sensing of environment 66.1 (1998),
ited on 03/16/2024).
pp. 1–16. DOI: 10.1016/S0034-4257(98)00031-5.
[316] Ting On Chan et al. “A novel reflectance transformation
[301] NASA. AERONET (Aerosol Robotic Network). Accessed:
and convolutional neural network framework for gener-
October 03, 2024. 2024. URL: https : / / aeronet .
ating bathymetric data for long rivers: A case study on
gsfc.nasa.gov/.
the Bei River in South China”. In: International Journal
[302] John Glasson and Riki Therivel. Introduction to envi- of Applied Earth Observation and Geoinformation 127
ronmental impact assessment. Routledge, 2013. ISBN: (Mar. 2024), p. 103682. ISSN: 15698432. DOI: 10.1016/
9781315881218. DOI: 10.4324/9781315881218. j.jag.2024.103682. (Visited on 03/16/2024).
[303] Mohammad Karimi Firozjaei et al. “Remotely sensed [317] Eyob Betru Wegayehu and Fiseha Behulu Muluneh. “Su-
urban surface ecological index (RSUSEI): an analytical per ensemble based streamflow simulation using multi-
framework for assessing the surface ecological status in source remote sensing and ground gauged rainfall data
urban environments”. In: Remote Sensing 12.12 (2020), fusion”. In: Heliyon 9.7 (July 2023), e17982. ISSN:
p. 2029. DOI: 10.3390/rs12122029. 24058440. DOI: 10.1016/j.heliyon.2023.e17982.
[304] Charles Spearman. The Proof and Measurement of As- (Visited on 03/16/2024).
sociation between Two Things. Appleton-Century-Crofts,
1961. DOI: 10.2307/1412159.
45
[318] Rami Al-Ruzouq et al. “Hybrid deep learning and re- [328] Ken Watanabe et al. “Improving accuracy and robustness
mote sensing for the delineation of artificial groundwa- of space-time image velocimetry (STIV) with deep learn-
ter recharge zones”. In: The Egyptian Journal of Remote ing”. In: Water 13.15 (2021), p. 2079. DOI: 10 . 3390 /
Sensing and Space Sciences 27.2 (June 2024), pp. 178– w13152079.
191. ISSN: 11109823. DOI: 10 . 1016 / j . ejrs . 2024 . [329] Eli Shusterman and Meir Feder. “Image compression via
02.006. (Visited on 03/16/2024). improved quadtree decomposition algorithms”. In: IEEE
[319] Hossein Sahour et al. “Identification of shallow ground- Transactions on Image Processing 3.2 (1994), pp. 207–
water in arid lands using multi-sensor remote sensing data 215. DOI: 10.1109/83.277901.
and machine learning algorithms”. In: Journal of Hydrol- [330] Andrew Howard et al. “Searching for MobileNetV3”. In:
ogy 614 (Nov. 2022), p. 128509. ISSN: 00221694. DOI: Proceedings of the IEEE/CVF International Conference
10 . 1016 / j . jhydrol . 2022 . 128509. (Visited on on Computer Vision (ICCV). Ed. by IEEE/CVF. 2019,
03/16/2024). pp. 1314–1324. DOI: 10.1109/ICCV.2019.00140.
[320] Jae Young Seo and Sang-Il Lee. “Predicting Changes in [331] Stuart K McFeeters. “The use of the Normalized Dif-
Spatiotemporal Groundwater Storage Through the Inte- ference Water Index (NDWI) in the delineation of open
gration of Multi-Satellite Data and Deep Learning Mod- water features”. In: International journal of remote sens-
els”. In: IEEE Access 9 (2021), pp. 157571–157583. ISSN: ing 17.7 (1996), pp. 1425–1432. DOI: 10 . 1080 /
2169-3536. DOI: 10 . 1109 / ACCESS . 2021 . 3130306. 01431169608948714.
(Visited on 03/16/2024).
[332] Mark J Van der Laan, Eric C Polley, and Alan E Hub-
[321] F. Sabzehee et al. “Enhancing spatial resolution of bard. “Super learner”. In: Statistical applications in ge-
GRACE-derived groundwater storage anomalies in Ur- netics and molecular biology 6.1 (2007). DOI: 10.2202/
mia catchment using machine learning downscaling meth- 1544-6115.1309.
ods”. In: Journal of Environmental Management 330
(Mar. 2023), p. 117180. ISSN: 03014797. DOI: 10.1016/ [333] Pierre Geurts, Damien Ernst, and Louis Wehenkel. “Ex-
j.jenvman.2022.117180. (Visited on 03/16/2024). tremely randomized trees”. In: Machine learning 63
(2006), pp. 3–42. DOI: 10.1007/s10994-006-6226-1.
[322] Wenlan Yang et al. “Monitoring multi-water quality of in-
ternationally important karst wetland through deep learn- [334] Viet-Ha Nhu et al. “A New Approach Based on Deep
ing, multi-sensor and multi-platform remote sensing im- Neural Networks and Multisource Geospatial Data for
ages: A case study of Guilin, China”. In: Ecological In- Spatial Prediction of Groundwater Spring Potential”. In:
dicators 154 (Oct. 2023), p. 110755. ISSN: 1470160X. IEEE Access 12 (2024), pp. 26344–26363. ISSN: 2169-
DOI : 10.1016/j.ecolind.2023.110755. (Visited on 3536. DOI : 10.1109/ACCESS.2024.3360337. (Visited
12/19/2023). on 03/16/2024).
[323] Taleb Vakili and Jamil Amanollahi. “Determination of op- [335] Suming Jin and Steven A Sader. “Comparison of time se-
tically inactive water quality variables using Landsat 8 ries tasseled cap wetness and the normalized difference
data: A case study in Geshlagh reservoir affected by agri- moisture index in detecting forest disturbances”. In: Re-
cultural land use”. In: Journal of Cleaner Production 247 mote sensing of Environment 94.3 (2005), pp. 364–372.
(Feb. 2020), p. 119134. ISSN: 09596526. DOI: 10.1016/ DOI : 10.1016/j.rse.2004.10.012.
j.jclepro.2019.119134. [336] Hanqiu Xu. “Modification of normalised difference wa-
[324] Dan Niu et al. “A Heterogeneous Spatiotemporal Atten- ter index (NDWI) to enhance open water features in re-
tion Fusion Prediction Network for Precipitation Now- motely sensed imagery”. In: International journal of re-
casting”. In: IEEE Journal of Selected Topics in Ap- mote sensing 27.14 (2006), pp. 3025–3033. DOI: 10 .
plied Earth Observations and Remote Sensing 16 (2023), 1080/01431160600589179.
pp. 8286–8296. ISSN: 1939-1404, 2151-1535. DOI: 10 . [337] Zhibin Sun et al. “Ensemble Learning via Higher Order
1109/JSTARS.2023.3310361. (Visited on 03/16/2024). Singular Value Decomposition for Integrating Data and
[325] Qizhao Jin et al. “SpatioTemporal Inference Network for Classifier Fusion in Water Quality Monitoring”. In: IEEE
Precipitation Nowcasting With Multimodal Fusion”. In: Journal of Selected Topics in Applied Earth Observations
IEEE Journal of Selected Topics in Applied Earth Obser- and Remote Sensing 14 (2021), pp. 3345–3360. ISSN:
vations and Remote Sensing 17 (2024), pp. 1299–1314. 1939-1404, 2151-1535. DOI: 10.1109/JSTARS.2021.
ISSN : 1939-1404, 2151-1535. DOI : 10 .1109 / JSTARS . 3055798. (Visited on 12/23/2023).
2023.3321963. (Visited on 03/16/2024). [338] Lieven De Lathauwer, Bart De Moor, and Joos Van-
[326] Chuyao Luo, Xutao Li, and Yunming Ye. “PFST-LSTM: dewalle. “A multilinear singular value decomposition”.
A SpatioTemporal LSTM Model With Pseudoflow Pre- In: SIAM journal on Matrix Analysis and Applica-
diction for Precipitation Nowcasting”. In: IEEE Journal tions 21.4 (2000), pp. 1253–1278. DOI: 10 . 1137 /
of Selected Topics in Applied Earth Observations and Re- S0895479896305696.
mote Sensing 14 (2021), pp. 843–857. ISSN: 1939-1404, [339] Lars Olof Bjorn et al. “A viewpoint: why chlorophyll a?”
2151-1535. DOI: 10 . 1109 / JSTARS . 2020 . 3040648. In: Photosynthesis research 99 (2009), pp. 85–98. DOI:
(Visited on 03/17/2024). 10.1007/s11120-008-9395-x.
[327] Md Touhidul Islam et al. “Mutual validation of re- [340] Christopher M Bishop. “Mixture density networks”. In:
mote hydraulic estimates and flow model simulations us- Aston University (1994).
ing UAV-borne LiDAR and deep learning-based imaging
techniques”. In: Results in Engineering 20 (Dec. 2023),
p. 101415. ISSN: 25901230. DOI: 10.1016/j.rineng.
2023.101415. (Visited on 03/16/2024).
46
[341] Sensen Wu et al. “Spatiotemporal assessments of nu- [353] Haozhe Zhang, Dan Nettleton, and Zhengyuan Zhu.
trients and water quality in coastal areas using remote “Regression-enhanced random forests”. In: arXiv preprint
sensing and a spatiotemporal deep learning model”. (2019). DOI: 10 . 48550 / arXiv . 1904 . 10416. arXiv:
In: International Journal of Applied Earth Observation 1904.10416. URL: https://arxiv.org/abs/1904.
and Geoinformation 112 (Aug. 2022), p. 102897. ISSN: 10416.
15698432. DOI: 10.1016/j.jag.2022.102897. (Vis- [354] Mark J Van der Laan. xView2 Dataset. Accessed: October
ited on 12/19/2023). 03, 2024. 2024. URL: https://xview2.org/dataset.
[342] Edson Filisbino Freire Da Silva et al. “A machine learn- [355] Ratiranjan Jena et al. “Earthquake risk assessment in NE
ing approach for monitoring Brazilian optical water types India using deep learning and geospatial analysis”. In:
using Sentinel-2 MSI”. In: Remote Sensing Applications: Geoscience Frontiers 12.3 (May 2021), p. 101110. ISSN:
Society and Environment 23 (Aug. 2021), p. 100577. 16749871. DOI: 10.1016/j.gsf.2020.11.007. (Vis-
ISSN : 23529385. DOI : 10 . 1016 / j . rsase . 2021 . ited on 12/20/2023).
100577. (Visited on 12/23/2023).
[356] Pan Xiong et al. “GNSS TEC-Based Earthquake Iono-
[343] R. J. Donaldson, Rosemary M. Dyer, and Michael J. spheric Perturbation Detection Using a Novel Deep
Kraus. “An Objective Evaluator of Techniques for Pre- Learning Framework”. In: IEEE Journal of Selected Top-
dicting Severe Weather Events”. In: Preprints, Ninth Con- ics in Applied Earth Observations and Remote Sensing
ference on Severe Local Storms, Norman, OK. Ed. by 15 (2022), pp. 4248–4263. ISSN: 1939-1404, 2151-1535.
American Meteorological Society. Vol. 321–326. Amer- DOI : 10 . 1109 / JSTARS . 2022 . 3175961. (Visited on
ican Meteorological Society, 1975. 12/20/2023).
[344] Junyoung Chung et al. “Empirical evaluation of gated re- [357] Amna Hafeez et al. “Possible Atmosphere and Iono-
current neural networks on sequence modeling”. In: arXiv spheric Anomalies of the 2019 Pakistan Earthquake Using
preprint (2014). DOI: 10 . 48550 / arXiv . 1412 . 3555. Statistical and Machine Learning Procedures on MODIS
URL: https://arxiv.org/abs/1412.3555. LST, GPS TEC, and GIM TEC”. In: IEEE Journal of Se-
[345] Andrew C Lorenc. “Analysis methods for numerical lected Topics in Applied Earth Observations and Remote
weather prediction”. In: Quarterly Journal of the Royal Sensing 14 (2021), pp. 11126–11133. ISSN: 1939-1404,
Meteorological Society 112.474 (1986), pp. 1177–1194. 2151-1535. DOI: 10 . 1109 / JSTARS . 2021 . 3119382.
DOI : 10.1002/qj.49711247414. (Visited on 12/20/2023).
[346] Zhaoxia Pu and Eugenia Kalnay. “Numerical weather [358] Yolanda Torres et al. “Integration of LiDAR and multi-
prediction basics: Models, numerical methods, and data spectral images for rapid exposure and earthquake vul-
assimilation”. In: Handbook of hydrometeorological en- nerability estimation. Application in Lorca, Spain”. In:
semble forecasting (2019), pp. 67–97. DOI: 10 . 1007 / International Journal of Applied Earth Observation and
978-3-642-40457-3_11-1. Geoinformation 81 (Sept. 2019), pp. 161–175. ISSN:
[347] Xingjian Shi et al. “Deep learning for precipitation now- 15698432. DOI: 10.1016/j.jag.2019.05.015. (Vis-
casting: A benchmark and a new model”. In: Advances in ited on 12/21/2023).
neural information processing systems 30 (2017). [359] Vivien Zahs et al. “Classification of structural build-
[348] Yunbo Wang et al. “Predrnn: Recurrent neural networks ing damage grades from multi-temporal photogrammet-
for predictive learning using spatiotemporal lstms”. In: ric point clouds using a machine learning model trained
Advances in neural information processing systems 30 on virtual laser scanning data”. In: International Journal
(2017). of Applied Earth Observation and Geoinformation 122
(Aug. 2023), p. 103406. ISSN: 15698432. DOI: 10.1016/
[349] Yichen Yao and Zhongjie Li. “CIKM AnalytiCup 2017: j.jag.2023.103406. (Visited on 12/21/2023).
short-term precipitation forecasting based on radar re-
flectivity images”. In: Proceedings of the Conference [360] Supattra Puttinaovarat and Paramate Horkaew. “Flood
on Information and Knowledge Management, Short-Term Forecasting System Based on Integrated Big and Crowd-
Quantitative Precipitation Forecasting Challenge, Singa- source Data by Using Machine Learning Techniques”. In:
pore. Ed. by CIKM. 2017, pp. 6–10. IEEE Access 8 (2020), pp. 5885–5905. ISSN: 2169-3536.
DOI : 10 . 1109 / ACCESS . 2019 . 2963819. (Visited on
[350] Dániel Erdélyi et al. “Predicting spatial distribution of 12/20/2023).
stable isotopes in precipitation by classical geostatistical-
and machine learning methods”. In: Journal of Hydrol- [361] Mohammed Moishin et al. “Designing Deep-Based
ogy 617 (Feb. 2023), p. 129129. ISSN: 00221694. DOI: Learning Flood Forecast Model With ConvLSTM Hy-
10 . 1016 / j . jhydrol . 2023 . 129129. (Visited on brid Algorithm”. In: IEEE Access 9 (2021), pp. 50982–
12/16/2023). 50993. ISSN: 2169-3536. DOI: 10.1109/ACCESS.2021.
3065939. (Visited on 12/20/2023).
[351] Tomislav Hengl, Gerard BM Heuvelink, and David G
Rossiter. “About regression-kriging: From equations to [362] Wenmei Li et al. “UNet Combined With Attention Mech-
case studies”. In: Computers and geosciences 33.10 anism Method for Extracting Flood Submerged Range”.
(2007), pp. 1301–1315. DOI: 10.1016/j.cageo.2007. In: IEEE Journal of Selected Topics in Applied Earth
05.001. Observations and Remote Sensing 15 (2022), pp. 6588–
6597. ISSN: 1939-1404, 2151-1535. DOI: 10 . 1109 /
[352] Aleksandar Sekulic et al. “Random forest spatial interpo- JSTARS.2022.3194375. (Visited on 12/20/2023).
lation”. In: Remote Sensing 12.10 (2020), p. 1687. DOI:
10.3390/rs12101687.

47
[363] Leila Hashemi-Beni and Asmamaw A. Gebrehiwot. [375] Ravinesh C Deo et al. “A real-time flood monitoring index
“Flood Extent Mapping: An Integrated Method Using based on daily effective precipitation and its application
Deep Learning and Region Growing Using UAV Opti- to Brisbane and Lockyer Valley flood events”. In: Water
cal Data”. In: IEEE Journal of Selected Topics in Ap- Resources Management 29 (2015), pp. 4075–4093. DOI:
plied Earth Observations and Remote Sensing 14 (2021), 10.1007/s11269-015-1046-3.
pp. 2127–2135. ISSN: 1939-1404, 2151-1535. DOI: 10 . [376] Fabio Montello, Edoardo Arnaudo, and Claudio Rossi.
1109/JSTARS.2021.3051873. (Visited on 12/20/2023). “MMFlood: A Multimodal Dataset for Flood Delineation
[364] Tejas Rathod et al. “Multipurpose deep learning-powered From Satellite Imagery”. In: IEEE Access 10 (2022),
UAV for forest fire prevention and emergency response”. pp. 96774–96787. ISSN: 2169-3536. DOI: 10 . 1109 /
In: HardwareX 16 (Dec. 2023), e00479. ISSN: 24680672. ACCESS.2022.3205419. (Visited on 12/20/2023).
DOI : 10 . 1016 / j . ohx . 2023 . e00479. (Visited on [377] National Oceanic and Atmospheric Administration
12/20/2023). (NOAA). National Oceanic and Atmospheric Administra-
[365] Van The Pham et al. “Classifying forest cover and map- tion (NOAA). Accessed: October 03, 2024. 2024. URL:
ping forest fire susceptibility in Dak Nong province, Viet- https://www.noaa.gov.
nam utilizing remote sensing and machine learning”. In: [378] J-SR Jang. “ANFIS: adaptive-network-based fuzzy infer-
Ecological Informatics 79 (Mar. 2024), p. 102392. ISSN: ence system”. In: IEEE transactions on systems, man, and
15749541. DOI: 10 . 1016 / j . ecoinf . 2023 . 102392. cybernetics 23.3 (1993), pp. 665–685. DOI: 10 . 1109 /
(Visited on 12/20/2023). 21.256541.
[366] Xiwen Chen et al. “Wildland Fire Detection and Monitor- [379] Dervis Karaboga and Bahriye Basturk. “A powerful and
ing Using a Drone-Collected RGB/IR Image Dataset”. In: efficient algorithm for numerical function optimization:
IEEE Access 10 (2022), pp. 121301–121317. ISSN: 2169- artificial bee colony (ABC) algorithm”. In: Journal of
3536. DOI: 10.1109/ACCESS.2022.3222805. (Visited global optimization 39 (2007), pp. 459–471.
on 12/20/2023).
[380] Bryce Hopkins Hopkins et al. “FLAME 2: Fire detection
[367] Ah Young Cho et al. “Burned Area Mapping Using and modeLing: Aerial Multi-spectral imagE dataset”. In:
Unitemporal PlanetScope Imagery With a Deep Learn- IEEE DataPort (2022). DOI: 10.21227/swyw-6j78.
ing Based Approach”. In: IEEE Journal of Selected Top-
[381] Alireza Shamsoshoara et al. “Aerial imagery pile burn de-
ics in Applied Earth Observations and Remote Sensing
tection using deep learning: The FLAME dataset”. In:
16 (2023), pp. 242–253. ISSN: 1939-1404, 2151-1535.
Computer Networks 193 (2021), p. 108001. DOI: 10 .
DOI : 10 . 1109 / JSTARS . 2022 . 3225070. (Visited on
1016/j.comnet.2021.108001.
12/20/2023).
[382] Andrew G Howard et al. “MobileNets: Efficient Convolu-
[368] Ali Akbar Siddique et al. “Sustainable collaboration: Fed-
tional Neural Networks for Mobile Vision Applications”.
erated learning for environmentally conscious forest fire
In: arXiv preprint (2017). DOI: 10.48550/arXiv.1704.
classification in Green Internet of Things (IoT)”. In: Inter-
04861. arXiv: 1704.04861.
net of Things 25 (Apr. 2024), p. 101013. ISSN: 25426605.
DOI : 10 . 1016 / j . iot . 2023 . 101013. (Visited on [383] Hamlyn G Jones and Robin A Vaughan. Remote sens-
12/20/2023). ing of vegetation: principles, techniques, and appli-
cations. Oxford University Press, USA, 2010. ISBN:
[369] George EP Box et al. Time series analysis: forecast-
9780199207794.
ing and control. John Wiley & Sons, 2015. ISBN: 0-13-
060774-6. [384] Xiaoli Wei. Wildfire Detection Dataset image. Accessed:
October 03, 2024. 2024. URL: https://www.kaggle.
[370] Judea Pearl. “Bayesian Networks: A Model of Self-
com/datasets/brsdincer/wildfire- detection-
Activated Memory for Evidential Reasoning”. In: Pro-
image-data.
ceedings of the 7th Conference of the Cognitive Science
Society. Ed. by Cognitive Science Society. University of [385] Siva R. Venna et al. “A Novel Data-Driven Model for
California, Irvine, CA, USA: Cognitive Science Society, Real-Time Influenza Forecasting”. In: IEEE Access 7
1985, pp. 15–17. (2019), pp. 7691–7701. ISSN: 2169-3536. DOI: 10.1109/
ACCESS.2018.2888585. (Visited on 03/31/2024).
[371] Judea Pearl. Probabilistic reasoning in intelligent sys-
tems: networks of plausible inference. Morgan kaufmann, [386] Hongping Hu et al. “Prediction of influenza-like illness
1988. ISBN: 978-0-08-051489-5. DOI: 10.1016/C2009- based on the improved artificial tree algorithm and artifi-
0-27609-4. cial neural network”. In: Scientific Reports 8.1 (Mar. 20,
2018), p. 4895. ISSN: 2045-2322. DOI: 10 . 1038 /
[372] J Hildebrand et al. “Simulating Lidar to create training
s41598-018-23075-1. (Visited on 03/31/2024).
data for machine learning on 3D point clouds”. In: ISPRS
Annals of the Photogrammetry, Remote Sensing and Spa- [387] Pi Guo et al. “Developing a dengue forecast model us-
tial Information Sciences 10 (2022), pp. 105–112. DOI: ing machine learning: A case study in China”. In: PLOS
10.5194/isprs-annals-X-4-W2-2022-105-2022. Neglected Tropical Diseases 11.10 (Oct. 16, 2017). Ed.
by Benjamin Althouse, e0005973. ISSN: 1935-2735. DOI:
[373] Lotfi Asker Zadeh. “Fuzzy sets”. In: Information and con-
10 . 1371 / journal . pntd . 0005973. (Visited on
trol 8.3 (1965), pp. 338–353.
03/31/2024).
[374] Seymour Geisser. “The predictive sample reuse method
[388] Adam Sadilek et al. “Machine-learned epidemiology:
with applications”. In: Journal of the American statistical
real-time detection of foodborne illness at scale”. In: npj
Association 70.350 (1975), pp. 320–328.
Digital Medicine 1.1 (Nov. 6, 2018), p. 36. ISSN: 2398-
6352. DOI: 10.1038/s41746-018-0045-1. (Visited on
03/31/2024).
48
[389] Yunxiang Zhao et al. “Predicting the transmission trend [401] Robert Tibshirani. “Regression shrinkage and selection
of respiratory viruses in new regions via geospatial sim- via the lasso”. In: Journal of the Royal Statistical Society
ilarity learning”. In: International Journal of Applied Series B: Statistical Methodology 58.1 (1996), pp. 267–
Earth Observation and Geoinformation 125 (Dec. 2023), 288. DOI: 10.1111/j.2517-6161.1996.tb02080.x.
p. 103559. ISSN: 15698432. DOI: 10 . 1016 / j . jag . [402] James G Kahn, Joshua S Yang, and James S Kahn. “‘Mo-
2023.103559. (Visited on 01/23/2024). bile’health needs and opportunities in developing coun-
[390] Cecilia Cordeiro Da Silva et al. “Covid-19 Dynamic Mon- tries”. In: Health affairs 29.2 (2010), pp. 252–258. DOI:
itoring and Real-Time Spatio-Temporal Forecasting”. In: 10.1377/hlthaff.2009.0965.
Frontiers in Public Health 9 (Apr. 8, 2021), p. 641253. [403] Santosh Kumar et al. “Center of Excellence for Mobile
ISSN : 2296-2565. DOI : 10.3389/fpubh.2021.641253.
Sensor Data-to-Knowledge (MD2K)”. In: IEEE Perva-
(Visited on 01/23/2024). sive Computing 16.2 (Apr. 2017), pp. 18–22. ISSN: 1536-
[391] Sofiane Atek et al. “A Geospatial Artificial Intelligence 1268. DOI: 10 . 1109 / MPRV . 2017 . 29. (Visited on
and satellite-based earth observation cognitive system in 03/31/2024).
response to COVID-19”. In: Acta Astronautica 197 (Aug. [404] Royal Netherlands Meteorological Institute (KNMI). Air
2022), pp. 323–335. ISSN: 00945765. DOI: 10.1016/j. Pollution - TEMIS. Accessed: October 03, 2024. 2024.
actaastro.2022.05.013. (Visited on 01/23/2024). URL : https : / / www . temis . nl / airpollution /
[392] Boxing Zhang et al. “COVID-19 contact tracking based absaai/.
on person reidentification and geospatial data”. In: Jour- [405] Yonghao Xu, Bo Du, and Liangpei Zhang. “Assessing
nal of King Saud University - Computer and Information the threat of adversarial examples on deep neural net-
Sciences 35.5 (May 2023), p. 101558. ISSN: 13191578. works for remote sensing scene classification: Attacks
DOI : 10.1016/j.jksuci.2023.101558. (Visited on
and defenses”. In: IEEE Transactions on Geoscience and
01/25/2024). Remote Sensing 59.2 (2020), pp. 1604–1617. DOI: 10 .
[393] Zhenyu Tan et al. “The Environmental Story During the 1109/TGRS.2020.2999962.
COVID-19 Lockdown: How Human Activities Affect [406] Jing Qiu et al. “The security of internet of vehicles net-
PM2.5 Concentration in China?” In: IEEE Geoscience work: Adversarial examples for trajectory mode detec-
and Remote Sensing Letters 19 (2022), pp. 1–5. ISSN: tion”. In: IEEE Network 35.5 (2021), pp. 279–283. DOI:
1545-598X, 1558-0571. DOI: 10 . 1109 / LGRS . 2020 . 10.1109/MNET.121.2000435.
3040435. (Visited on 01/23/2024).
[407] Li Chen et al. “Lie to Me: A Soft Threshold Defense
[394] Yu Zhan et al. “Spatiotemporal prediction of continu- Method for Adversarial Examples of Remote Sensing Im-
ous daily PM2.5 concentrations across China using a ages”. In: IEEE Geoscience and Remote Sensing Let-
spatially explicit machine learning algorithm”. In: Atmo- ters 19 (2022), pp. 1–5. ISSN: 1545-598X, 1558-0571.
spheric Environment 155 (Apr. 2017), pp. 129–139. ISSN: DOI : 10 . 1109 / LGRS . 2021 . 3096244. (Visited on
13522310. DOI: 10.1016/j.atmosenv.2017.02.023. 03/02/2024).
(Visited on 03/31/2024).
[408] Abdul Jhummarwala, MB Potdar, and Prashant Chauhan.
[395] Qian Di et al. “Assessing PM 2.5 Exposures with High “Parallel and distributed GIS for processing geo-data: an
Spatiotemporal Resolution across the Continental United overview”. In: International Journal of Computer Appli-
States”. In: Environmental Science & Technology 50.9 cations 106.16 (2014), pp. 9–16. DOI: 10.5120/18602-
(May 3, 2016), pp. 4712–4721. ISSN: 0013-936X, 1520- 9881.
5851. DOI: 10 . 1021 / acs . est . 5b06121. (Visited on
03/31/2024). [409] Jose E. Adsuara et al. “Discovering Differential Equations
from Earth Observation Data”. In: IGARSS 2020 - 2020
[396] Centers for Disease Control and Prevention. Centers for IEEE International Geoscience and Remote Sensing Sym-
Disease Control and Prevention Website. Accessed: Oc- posium. IGARSS 2020 - 2020 IEEE International Geo-
tober 03, 2024. 2024. URL: https://www.cdc.gov/. science and Remote Sensing Symposium. Ed. by IEEE
[397] Andrea Freyer Dugas et al. “Influenza forecasting with Geoscience and Remote Sensing Society. Waikoloa, HI,
Google flu trends”. In: PloS one 8.2 (2013), e56176. DOI: USA: IEEE, 2020, pp. 3999–4002. ISBN: 978-1-72816-
10.1371/journal.pone.0056176. 374-1. DOI: 10.1109/IGARSS39084.2020.9324639.
[398] Fred S. Lu et al. “Improved state-level influenza nowcast- [410] Alex Levering et al. “Interpretable Scenicness from
ing in the United States leveraging Internet-based data and Sentinel-2 Imagery”. In: IGARSS 2020 - 2020 IEEE In-
network approaches”. In: Nature Communications 10.1 ternational Geoscience and Remote Sensing Symposium.
(Jan. 11, 2019), p. 147. ISSN: 2041-1723. DOI: 10.1038/ IGARSS 2020 - 2020 IEEE International Geoscience and
s41467-018-08082-0. (Visited on 03/31/2024). Remote Sensing Symposium. Ed. by IEEE Geoscience
[399] Centers for Disease Control and Prevention. Weekly US and Remote Sensing Society. Waikoloa, HI, USA: IEEE,
Influenza Surveillance Report. Accessed: October 03, Sept. 26, 2020, pp. 3983–3986. ISBN: 978-1-72816-374-
2024. Centers for Disease Control and Prevention, 2024. 1. DOI: 10.1109/IGARSS39084.2020.9323706. (Vis-
URL: https://www.cdc.gov/flu/weekly/usmap. ited on 01/04/2024).
htm. [411] Chia-Yu Hsu and Wenwen Li. “Explainable GeoAI: can
[400] Jerome H Friedman. “Greedy function approximation: saliency maps help interpret artificial intelligence’s learn-
a gradient boosting machine”. In: Annals of statis- ing process? An empirical study on natural feature detec-
tics (2001), pp. 1189–1232. DOI: 10 . 1214 / aos / tion”. In: International Journal of Geographical Informa-
1013203451. tion Science 37.5 (2023), pp. 963–987. DOI: 10.1080/
13658816.2023.2191256.

49
[412] Ximeng Cheng et al. “Explainability in GeoAI”. In: [417] SpatioTemporal Asset Catalog. STAC Specification. Ac-
Handbook of Geospatial Artificial Intelligence. Ed. by cessed: October 03, 2024. 2023. URL: https : / / www .
Editors of Handbook of Geospatial Artificial Intelligence. stacspec.org/.
CRC Press, 2024, pp. 177–200. ISBN: 9781003308423. [418] GeoRSS. Official Website of GeoRSS. Accessed: October
[413] Rob Van Kranenburg. The Internet of Things. A critique of 03, 2024. 2023. URL: http://www.georss.org/.
ambient technology and the all-seeing network of RFID. [419] Open Geospatial Consortium. Official Website of the
Vol. 2. Institute of Network Cultures, 2007. ISBN: 978-90- Open Geospatial Consortium. Accessed: October 03,
78146-06-3. 2024. 2023. URL: https://www.ogc.org/.
[414] Uichin Lee et al. “Mobeyes: smart mobs for urban mon-
[420] Constituency Parsing. “Speech and language processing”.
itoring with a vehicular sensor network”. In: IEEE Wire-
In: Power Point Slides (2009).
less Communications 13.5 (2006), pp. 52–57. DOI: 10 .
1109/WC-M.2006.250358. [421] Christel Tartini-Chappuis et al. “Multi-task prompt-
RSVQA to explicitly count objects on aerial images”. In:
[415] Dan Puiu et al. “Citypulse: Large scale data analytics
British Machine Vision Conference (BMVC) workshops.
framework for smart cities”. In: IEEE Access 4 (2016),
Ed. by BMVC. 2023.
pp. 1086–1108. DOI: 10.1109/ACCESS.2016.2541999.
[422] Haonan Guo et al. Remote Sensing ChatGPT: Solving Re-
[416] S Mirdula and M Roopa. “MUD enabled deep learning
mote Sensing Tasks with ChatGPT and Visual Models.
framework for anomaly detection in IoT integrated smart
Accessed: October 03, 2024. Jan. 17, 2024. DOI: 10 .
building”. In: e-Prime-Advances in Electrical Engineer-
48550/arXiv.2401.09083. (Visited on 03/02/2024).
ing, Electronics and Energy 5 (2023), p. 100186. DOI:
10.1016/j.prime.2023.100186.

50