Uncorrected Proof

1 © IWA Publishing 2020 Water Science & Technology | in press | 2020

A comprehensive review of deep learning applications

in hydrology and water resources
Muhammed Sit, Bekir Z. Demiray, Zhongrun Xiang, Gregory J. Ewing,
Yusuf Sermet and Ibrahim Demir

ABSTRACT

The global volume of digital data is expected to reach 175 zettabytes by 2025. The volume, variety, Muhammed Sit (corresponding author)
Interdisciplinary Graduate Program in Informatics,
and velocity of water-related data are increasing due to large-scale sensor networks and increased University of Iowa,
United States
attention to topics such as disaster response, water resources management, and climate change. and
IIHR - Hydroscience & Engineering,
Combined with the growing availability of computational resources and popularity of deep learning,
The University of Iowa,
these data are transformed into actionable and practical knowledge, revolutionizing the water 100 C. Maxwell Stanley Hydraulics Laboratory,
Iowa City,
industry. In this article, a systematic review of literature is conducted to identify existing research Iowa
52242-1585
which incorporates deep learning methods in the water sector, with regard to monitoring, E-mail: muhammed-sit@uiowa.edu

management, governance and communication of water resources. The study provides a

Bekir Z. Demiray
comprehensive review of state-of-the-art deep learning approaches used in the water industry for Department of Computer Science,
University of Iowa,
generation, prediction, enhancement, and classification tasks, and serves as a guide for how to United States

utilize available deep learning methods for future water resources challenges. Key issues and Zhongrun Xiang
Gregory J. Ewing
challenges in the application of these techniques in the water domain are discussed, including the Ibrahim Demir
ethics of these technologies for decision-making in water resources management and governance. Department of Civil and Environmental
Engineering,
Finally, we provide recommendations and future directions for the application of deep learning University of Iowa,
United States
models in hydrology and water resources.
Yusuf Sermet
Key words | artificial intelligence, deep learning, hydroscience, machine learning, review, water
Department of Electrical and Computer
Engineering,
University of Iowa,
United States
HIGHLIGHTS

• A comprehensive review of deep learning applications in hydrology and water

resources fields is conducted.
• Summaries of papers that employ deep learning for hydrologic modeling is presented
along with a concise deep learning synopsis.
• Key issues within the water-related deep learning application literature are reported
• Future directions for research in the intersection of deep learning and hydroscience is
presented.

LIST OF ACRONYMS

AE Autoencoder BP BackPropagation
AI Artificial Intelligence BPNN BackPropagation Neural Networks
ANN Artificial Neural Networks CLSTM, Convolutional LSTM
ARIMA AutoRegressive Integrated Moving ConvLSTM
Average CNN, ConvNet Convolutional Neural Network
BMP Best Management Practices CSO Combined Sewer Overflow
doi: 10.2166/wst.2020.369

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
2 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

DBM Deep Boltzman Machines RBM Restricted Boltzman Machine

DBN Deep Belief Network RGB Red Green Blue
DEM Digital Elevation Model RNN Recurrent Neural Network
DL Deep Learning ReLU Rectified Linear Unit
DNN Deep Neural Network SAE Stack AutoEncoder
DQN Deep Q-Network SD Statistical Downscaling
DWT Discrete Wavelet Transform SDAE Stacked Denoising AutoEncoders
EC Electrical Conductivity SRGAN Super Resolution Generative Adversarial
EEMD Empirical Mode Decomposition Networks
EFDC Environmental Fluid Dynamics SSIM Structural Similarity
ELM Extreme Learning Machines SST Sea Surface Temperature
ENN Elman Neural Network SVM Support Vector Machines
FCN Fully Convolutional Network SVR Support Vector Regression
FFNN Feed-Forward Neural Network SWAT Soil and Water Assessment Tool
GAN Generative Adversarial Networks USGS United States Geological Survey
GIS Geographic Information System VAE Variational AutoEncoders
GP Gaussian Process WLSTM Wavelet-LSTM
GPROF Goddard Profiling Algorithm
GPU Graphics Processing Unit
GRACE Gravity Recovery and Climate
Experiment INTRODUCTION
GRU Gated Recurrent Unit
IMF Intrinsic Mode Function The global volume of digital data is expected to reach 175
zettabytes by 2025 (Reinsel et al. ). Large-scale sensor
KNN K-Nearest Neighbors
networks as well as the increased awareness of climate
LSTM Long Short-Term Memory
change, water resources management, and the monitoring
LiDAR Light Detection and Ranging
of water-related hazards led to the substantial growth of
MAP Mean Areal Precipitation
the volume, variety, and velocity of water-related data
MAPLE McGill Algorithm for Precipitation now- (Weber et al. ; Sit et al. a, b). Modern data col-
casting by Lagrangian Extrapolation lection techniques, including satellite hydrology, internet of
MBR Membrane Bioreactor things for on-site measurements (Kruger et al. ), and
ML Machine Learning crowdsourcing tools (Sermet et al. b), has revolutio-
MLP Multi-Layer Perceptron nized the water science and industry as approached by the
MSE Mean Squared Error government, academia, and private sector (Krajewski et al.
NAR Nonlinear Autoregressive ). The effective utilization of vast water data holds the
NARX Nonlinear Autoregressive eXogenous key for long-term sustainability and resilience and presents
NDVI Normalized Difference Vegetation Index opportunities to transform water governance for the upcom-
ing decades (Grossman et al. ). In the hydrological
NN Neural Network
domain, multivariate analysis relying on extensive and
NOAH National Oceanic and Atmospheric
semantically-connected data resources is required to gener-
Administration
ate actionable knowledge and produce realistic and
ORP Oxidation-Reduction Potential
beneficial solutions to water challenges facing communities
PDE Partial Differential Equation (Carson et al. ; Jadidoleslam et al. ). However, the
PLSR Partial Least Square Regression inaccessible, unstructured, nonstandardized, and incompati-
POD Proper Orthogonal Decomposition ble nature of the data makes optimized data models (Demir
PaaS Platform as a Service & Szczepanek ) and smarter analytics approaches a
QPF Quantitative Precipitation Forecasts necessity (Sermet & Demir a).

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
3 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Computerized methods to create an understanding of This paper presents a systematic review of applications
hydrological phenomena are based on various modeling of deep learning within the scope of the hydrological
strategies, which simplify a hydrological system to simulate domain. The literature has been thoroughly examined to
its behavior (Antonetti & Zappa ). Physical models identify the use cases of deep learning in the subfields of
aim to achieve this goal by specifically designing complex the water sector including flooding, weather, land use and
simulations that are powered by mathematical and numeric soil, water quality, surface water, water resources, and
specifications of conceptualized physical characteristics groundwater. Each study has been evaluated to extract infor-
(Jaiswal et al. ). However, hydrological systems, as is mation that is scientifically relevant to assess the study’s
the case with other natural systems, are inherently hetero- contribution and reproducibility including the hydrological
geneous (Marçais & De Dreuzy ) as opposed to less tasks that were taken on to be approached by deep learning
complex human-made systems with defined rules. Therefore along with the utilized network architectures, datasets, soft-
physical models, though are deterministic and reliable, do ware tools and frameworks, licenses, and deep learning
not always perform and scale well due to their intrinsic limit- practices and algorithms. The paper explores modern deep
ations (Islam ). As an alternative, statistical models have learning networks from the lens of the hydrological domain
been employed to make use of the comprehensive set of avail- to investigate the shortcomings and challenges of the status
able hydrological, environmental, and geophysical data quo and to highlight the opportunities and future directions
(Evora & Coulibaly ). These approaches assume mini- to serve as a guide to researchers, professionals, and govern-
mum awareness of the underlying mechanism and receive mental organizations in the water sector.
their strength by eliciting useful information and patterns The major contributions of this paper can be summar-
from the available data through statistical analysis (McCuen ized as follows. Though there are various configurations of
). However, they have displayed shortcomings in terms artificial neural networks optimized for various data types
of accuracy and certainty, and also require excessive compu- and use cases, it is challenging to reduce a real-life hydrolo-
tational power (Ardabili et al. ; Agliamzanov et al. ). gical task to a certain predefined approach given the depth
Recent developments in artificial intelligence and and complexity of the tasks as well as the diversity of
graphical processor units (GPU) have paved the way for networks. The methodologies that need to be employed
deep learning (DL), a pioneering approach that is fueled while developing deep learning-powered solutions in hydrol-
by multilayer artificial neural networks (LeCun et al. ). ogy are not standardized in terms of data quality and
Deep learning provides a black-box method to learn from preparation, execution, validation, and documentation.
complex and high-dimensional data to infer robust and scal- Furthermore, the strength, usability, and reliability of a
able insights while minimizing the degree to which manual model lie on clearly set descriptions and procedures for
labor is needed (Sengupta et al. ). One feature which deterministic reproducibility, given the variety of develop-
separates deep learning from its superset machine learning ment frameworks as well as the application areas. To the
(ML) is the use of multilayer models which leads to a best of our knowledge, there has not been a thorough inves-
higher-level representation of the underlying data sources tigation of systematically approaching water challenges with
(Saba et al. ). Furthermore, deep learning is capable of deep learning. Thus, this paper serves as a meticulous guide
extracting substantial features without being explicitly for the stakeholders of the hydrology domain to advance the
instructed, and thus, is more immune to raw and noisy water industry with intelligent systems that are revolutio-
data (Sahiner et al. ). Successful implementations of nized by multi-faceted data (Sermet et al. a).
deep learning permeate numerous domains and industries The remainder of this article is organized as follows.
including medical imaging (Haskins et al. ), healthcare Literature Review section provides the review methodology
(Esteva et al. ), finance (Heaton et al. ), geophysical followed by a comprehensive literature review of deep learn-
sciences (Shen ), remote sensing (Ma et al. ), and ing applications in the water domain. Descriptions of deep
hydrology. Due to its significant adoption rate and potential learning concepts, tasks, and architectures are described to
to be applicable to any domain which encompasses pro- summarize the available methodology for use by the hydro-
blems that can be expressed as control systems, numerous logical community. Results section presents a detailed
open-source and for-profit software tools, educational summary and analysis of reviewed papers grouped by their
resources, and generalized algorithms have been made avail- application area. Key Issues and Challenges section high-
able for use, opening up countless paths to advance lights the key issues and challenges facing the stakeholders
hydrological studies. utilizing deep learning in the water domain with respect to

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
4 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

technical limitations as well as ethical considerations. abstracts were included in the first list of articles gathered.
Recommendations and Conclusions section outlines a This time interval is primarily chosen based on our initial lit-
vision entailing the adoption of prominent and deep learn- erature search and availability of deep learning application
ing-powered technologies to solve the water challenges of papers enough to create a comprehensive review and
the future and then, concludes the paper with a concise sum- curate insights within the water domain between 2018 and
mary of findings. 2020. There were also other review articles partially cover-
ing the water domain and timeline (Shen ). After
gathering the initial list totaling 1,515 publications, each of
LITERATURE REVIEW them was briefly reviewed to determine whether they were
in alignment with the scope of this study. All publications
This section starts with a detailed description of the litera- that are not research papers were excluded, namely vision
ture search methodology in the first subsection Review papers, editorials and review papers. From the initial list
Methodology and then presents the information extracted of 1,515 publications, 315 remained after this filtering step.
from each reviewed manuscript. The subsection Deep These publications were filtered further to keep publications
Learning gives a brief overview of deep learning history, that met certain technical criteria. This step eliminated all
describes various neural network architectures, and elabor- publications that did not involve some form of deep artificial
ates on different machine learning task types. At the end neural network in their pipeline of work.
of this section we share the summary of the literature as After this step 129 publications remained and were
figures to provide an understanding of this review and a included in our comprehensive review. The comprehensive
table of all the papers reviewed. review process consisted of manually reviewing the papers
one by one to extract specific publication features, including:
Architecture, Framework/Library/Programming Language,
REVIEW METHODOLOGY Dataset, Source Code Sharing, Reproducibility, Subfield:
Deep Learning, Subfield: Environment, Summary. Each of
A systematic literature search on water domain was the feature categories are described below:
employed for this review. Web of Science, Scopus, Springer
Link, Wiley Online Library and The International Water
Association Publishing Online were used as the databases • Architecture – The type of deep neural network architec-
and the keywords included ‘deep neural network’, ‘deep ture(s) employed in the study. This could be simply
neural networks’, ‘deep learning’, ‘lstm’, ‘long short term Artificial Neural Network (ANN) or more complex archi-
memory’, ‘cnn’, ‘convolutional’, ‘gan’, ‘generative adversar- tectures like Generative Adversarial Networks (GAN) or
ial’, ‘rnn’, ‘recurrent neural’, ‘gru’ and ‘gated recurrent’. Autoencoders (VAE).
Hydroscience and water resources specific keywords were • Framework/Library – This column serves as a survey
not included in the search queries because of the cardinality within the field to understand the programming language
of domain-specific keywords. The cardinality is a result of and numeric computation library choices of researchers.
many application fields within these domains and also • Dataset – Whether the dataset(s) used in the study were
many interchangeable and synonymous words used to collected specifically for the study, acquired from an auth-
refer the same terms within different scientific communities ority resource, or previously existing standalone.
around the globe. Hydrology and water resources related • Source Code Sharing – A boolean field indicating if the
papers are filtered at the journal level which provided a code of the study is open sourced and accessible by the
comprehensive coverage compared to keyword level filter- public.
ing. After limiting the search with these keywords in • Reproducibility – A boolean field indicating if the results
publication title, abstract or keywords, an additional exclu- of the study could be reproduced just by using the infor-
sion criterion was applied through each database’s mation provided in the manuscript.
categorization system if applicable only to include the publi- • Subfield: Deep Learning – A classification of the
cations within the environmental fields. Also, searches were machine learning task tackled in the paper. This field
limited to only include journal publications. All articles pub- uses one of the following values; Regression, Classifi-
lished in 2018, 2019 and 2020 up until the end of March cation, Sequence Prediction, Matrix Prediction,
containing these keywords in their titles, keyword fields or Unsupervised Learning and, Reinforcement Learning.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
5 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Details of these are given in the next subsection where we In modern neural networks the common (Goodfellow
describe deep learning practices. et al. ) and recommended (Nair & Hinton ) acti-
• Subfield: Environment – A classification of the task car- vation function is the Rectified Linear Unit (ReLU)
ried out in an environmental field. This field uses one of function which introduces non-linearity to the network. A
the following values; Flood, Groundwater, Land Usage hidden layer that applies ReLU to the input could be
and Soil, Surface Water, Water Quality, Water Resources referred as ReLU layer or an activation layer using ReLU.
Management, Weather and, Others. Others include ANNs are generally trained using the backpropagation
papers within the environmental field but do not exactly (BP) algorithm (Rumelhart et al. ) to fit the network to
fit with other subfields we included here. the set of input-output pairs using a loss function. A loss
• Summary – The brief summary of the study. function or a cost function is used to determine how success-
ful the mapping is done by a model. An easy and go-to loss
These data for each publication reviewed herein are
shared later in this section with figures and a table. Techni- function is mean squared error (MSE) which computes the
cal summaries of the papers reviewed are given in the difference between an output of the ANN and the ground
truth for each input-output sample, squares the difference
Results section. Conclusions drawn from the acquired data
are shared in the Key Issues and Challenges section. to avoid negative values and computes the mean error of
all the samples. BP trains an ANN using the loss function
by computing the gradient of the loss function with respect
DEEP LEARNING to the weights between neurons and updates the weights to
better fit the data samples in order to minimize the loss
Deep Learning is a subfield of Machine Learning where a function.
long-known algorithm, a neural network (NN) or an artifi- ANNs are not a new concept, considering primitive ver-
cial neural network (ANN) or a deep neural network sions were around in the 1940s (Goodfellow et al. ).
(DNN) or a feed-forward neural network (FFNN), referring Instead, ANNs attracted attention of researchers from var-
to the same algorithm, is used to map features into an output ious scientific disciplines when it became clear that they
or a set of outputs. Formed by an input layer, intermediate are extremely powerful in capturing representations in the
hidden layers and an output layer, ANNs present an efficient data and with advances in graphics processing units
way to learn linear and non-linear relationships between (GPUs) which enable extremely fast matrix operations
input and output pairs. Neural networks, when formed by (Goodfellow et al. ). In this way, a neural network archi-
many stacked layers, can represent complex features in tecture that previously was infeasible to utilize due to time
later layers by using simpler representations formed by ear- complexity as the number of hidden layers increased,
lier layers in the network (Goodfellow et al. ). Each could be used in training on complex datasets learning
layer within an ANN comprises at least a neuron. ANN is representations.
a network of these neurons connected to each other with For each of the reviewed papers we identify the machine
some weights and these neurons run specific functions, learning subfield, which correspond to task types. Task types
namely activation functions, mapping it’s input to an are defined by the output form of a network. We consider
output. Stacked on top of each other, the series of functions the following as distinct task types: Regression, Classifi-
runs over the input of the network, translates the input to cation, Sequence Prediction, Matrix Prediction,
the output in the output layer. Typically, each neuron Unsupervised Learning, and Reinforcement Learning.
within a layer runs the same activation function and type These task types are briefly discussed below:
of the layer is determined by this activation function. Net-
work type is determined by the combination of layers used • Regression – A machine learning task that predicts a con-
and how neurons are connected to each other within and tinuous value. Forecasting the next measurement from a
between layers. The quintessential form of an ANN is the series of measurements would be an example of a
multilayer perceptron (MLP). An MLP contains at least a regression task.
single hidden layer while each neuron within the network • Classification – When the aim of a model is to predict
is connected to every neuron within the next layer. This whether the set of inputs can be categorized into some
architecture forms a fully connected neural network. classes, the task is a classification task. The number of
An activation function in a typical MLP multiplies the classes are not limited. Predicting whether the next day
input by a weight and outputs it to the next neuron in line. will be a dry or wet one is a classification task.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
6 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

• Sequence Prediction – Regression of a sequence of output along with linear correlations, one needs to intro-
numeric values or a vector. Forecasting the next 24 duce non-linearity to the network via an activation layer.
hours of measurements for a stream sensor is a sequence Typically, that is done by using the Rectified Linear Unit
prediction task. (ReLU) as the activation function following the convolution
• Matrix Prediction – Regression of a matrix of numeric layer. Another common layer used within a CNN is a pool-
values. Forecasting the next precipitation measurements ing layer. A pooling layer is used to reduce the size of the
for a rectangular region would be a matrix prediction input while keeping the positional knowledge intact. A fre-
task. Each of the values in the predicted matrix would quently used pooling method within CNN literature is
be the precipitation value of a subregion within the Max Pooling (Zhou & Chellappa ). This sample-based
actual region. discretization moves the most important learnt features to
• Unsupervised Learning – A learning task applying com- subsequent layers while reducing the size. Consequently,
petitive learning instead of error correction learning like CNNs make good architectures for deep learning tasks
previous task types. Decreasing dimension of a high- with images or image-like objects as inputs. This ability of
dimensional hydrological input data to lower-dimension CNNs makes way for various breakthroughs in the fields
visualizable data would be an example. of object detection, super-resolution, image classification,
• Reinforcement Learning – A learning task where the and computer vision.
output is unspecified. In a reinforcement learning task,
the algorithm tries to find the optimal solution for any
given input using a reward/penalty policy and a try- GENERATIVE ADVERSARIAL NETWORKS (GANS)
error mechanism. An example of a reinforcement learn-
ing task is an artificial intelligence (AI) model that GANs (Goodfellow et al. ) consist of two seemingly sep-
learns when to release water from a dam. arate CNNs working in unison and competing in a min-max
game. One of these CNNs, the generator, aims to generate
fake examples out of a dataset while the other, discrimina-
ARCHITECTURES tor, aims to reveal whether its input is fake or not. Since
they try to beat each other, it causes them to get better
This subsection summarizes cornerstone neural network over time in both generating fake outputs and discriminating
architectures used by papers reviewed in this study. Also, fake from real.
some ANN concepts are briefly discussed. For further GANs are initially used as generative models, as in ran-
understanding of these architectures, we refer readers to domly generating new samples from a dataset to appear as if
the cited works. they are from the originating dataset when visualized
(Gautam et al. ). They achieve this goal by mapping
random noise to real samples from the given dataset, and
CONVOLUTIONAL NEURAL NETWORKS (CNNS) then they generate new instances from new random noise
tensors. Despite their success at generation, GANs are also
A CNN (LeCun ) or a ConvNet comprises at least one capable of learning translation tasks such as super-resol-
convolutional layer, which methodologically expects a 3D ution (Demiray et al. ) or image to image translation
tensor as the input. A convolutional layer applies multiple (Isola et al. ).
cascaded convolution kernels to extract intricate knowledge
from the input. For example, a CNN for an RGB image
tensor with a shape of image width x image height × 3 RECURRENT NEURAL NETWORKS (RNNS)
would have a convolutional layer which applies 3 different
convolution operations with 3 separate kernels to each of RNNs (Pollack ) are a type of artificial neural network
the color channel matrices. Using a convolution kernel that includes a recurrent layer. The difference of a recurrent
matrix, a convolutional layer that processes an image as layer from a regular fully-connected hidden layer is that
such can extract 2D positional information from images, neurons within a recurrent layer could be connected to
such as understanding objects that are close to each other. each other as well. In other words, the output of a neuron
To make a neural network with a convolutional layer is conveyed both to the neuron(s) within the next layer
cognize non-linear correlations between the input and and to the next neuron within the same layer. Using this

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
7 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

mechanism, RNNs can carry information learned within a such as long paragraphs, as it is common to encounter
neuron to the next neuron in the same layer. the vanishing gradient problem while training (Bengio
This procedure becomes convenient when the data to et al. ). With the vanishing gradient problem, the gradi-
be fed to the network is in sequential nature such as a ents of the loss function get extremely high in some cases
time-series data or a text. When training a model over a during training and consequently make the training pro-
data sample like a text to extract the meaning, most of cess and the trained network deficient (Goodfellow et al.
the time, the beginning of the text could change the mean- ). More complex RNN implementations like Long
ing that is to be extracted from the end of the text. RNNs Short-Term Memory (LSTM) (Hochreiter & Schmidhuber
aim to keep the information gained from earlier parts of ) Networks or Gated Recurrent Unit (GRU) Networks
a data sample in the memory and move it to the later (Cho et al. ) solve this problem but have greater com-
parts of the same data sample to ensure better knowledge putational complexity. Various RNN structures can be
discovery (Goodfellow et al. ). A simple RNN used for tasks that somewhat rely on sequential under-
implementation lacks the practicality in long sequences, standing of datasets such as language modeling

Figure 1 | A densely connected artificial neural network architecture.

Figure 2 | A basic convolutional neural network structure for image classification.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
8 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

(Sundermeyer et al. ), text classification (Yin et al. ; GATED RECURRENT UNIT (GRU) NETWORKS
Sit et al. a, b) and time-series forecasting (Sit &
Demir ; Xiang & Demir ). Although LSTM Networks, most of the time, solve the van-
ishing gradients problem and helped many breakthroughs
within the fields of Natural Language Processing and time-
series prediction, their time complexity emerges as a down-
LONG SHORT-TERM MEMORY (LSTM) NETWORKS side. GRU (Cho et al. ) networks reduce the complexity
while keeping the efficacy intact. Similar to a simple RNN
LSTM (Hochreiter & Schmidhuber ) Networks are neuron, a GRU neuron produces only one output after a
developed for longer short term memory life over the series of computations and uses the same output to convey
input, paving the way for more efficient but more resource important features learnt to both the next layer and the
intensive training over datasets consisting of sequential next neuron within the same layer.
samples. Instead of an activation function producing one
output and carrying the output to immediate neurons both
in the next layer and the same layer, LSTM neurons produce
two different values yielded by a series of activations and NONLINEAR AUTOREGRESSIVE (NAR) MODELS
operations. While both outputs are kept within the LSTM
layer to keep track of things learnt over the past part of A NAR model is not necessarily a neural network model but
the sequence, one of the outputs is transferred to the next a model that is used for time-series prediction, taking
layer (Figure 5). into account both current and previous samples from a

Figure 3 | Overview of generative adversarial networks.

Figure 4 | Connections of a Recurrent layer.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
9 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Figure 5 | Computation-wise comparison of RNN, LSTM and GRU nodes.

Figure 6 | An Elman Network architecture with two input and two output neurons.

time-series to map input sequence to outputs. A NAR model ELMAN NETWORK (ENN)
needs a nonlinear function such as a polynomial function or
a neural network to perform training. If a neural network is An Elman Network (Elman ) is yet another RNN
used, a NAR network would classify as an RNN based on implementation that has three layers, only one being a
the fact that it utilizes sequential complexion of the given hidden layer. The hidden layer of the Elman Network is con-
input. Papers reviewed within this study that employ NAR, nected to a set of neurons called context units. In each
thus, implement a neural network as the function in their iteration after the first one over the network, the state of
proposed models. There are many NAR variations and one the hidden layer is copied into the context units. Thus, the
that deserves mention, due to cardinality of papers reviewed state of the network for the previous sample in the data
in this study which employ it, is the Nonlinear Autoregres- stream is kept in the network each time to be used in next
sive Exogenous Model (NARX) (Lin et al. ). NARX is iterations. An Elman Network can train over sequential
an RNN implementation that takes advantage of exogenous datasets better than a regular ANN due to this mechanism
inputs, hence the name. acting like a memory.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
10 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Figure 7 | Visualization of an Autoencoder network.

AUTOENCODERS (AE) set of fully-connected layers in order to perform a classifi-

cation task. Formed by stacked RBMs, DBNs, similarly to
Autoencoders (Rumelhart et al. ) are neural networks RBMs, could also be used to tackle many types of tasks
that are used to reduce the dimensionality of datasets. by training unsupervised beforehand. The bipartite connec-
They are implemented in an unsupervised fashion to gener- tions of RBMs are altered when they are stacked onto each
ate only a representation of the dataset within their hidden other to form a DBN (Figure 8). Only the top two layers of
layer neurons, also called the latent vector. Taking the a DBN have bipartite connections while the rest of the
same set of values for both input and output of the network, layers have one-way connections with each other. Though
an AE learns to reduce a dataset into a representation state RBMs are useful for many tasks, their use in research is
and additionally learns how to reconstruct the data sample decreasing in the machine learning field as researchers
to its original form from the learned representation. adopt newer architectures that could be utilized in the
same fashion such as AE, DBNs and GANs.

RESTRICTED BOLTZMANN MACHINES (RBMS) AND

DEEP BELIEF NETWORKS (DBNS) EXTREME LEARNING MACHINES (ELM)

RBMs (Hinton ) present two-layer stochastic genera- Extreme Learning Machines (Huang et al. ) are three-
tive models that are in the shape of bipartite graphs. layer neural networks in which the weights connecting to
RBMs form the building block for deep belief networks the second layer from the input layer are randomized, and
(DBN) (Hinton ), but they can also be used as standa- the weights connecting to the third are trained. ELMs are
lone models. RBMs were initially used in unsupervised designed to produce better generalization performance
tasks, they also enable the user to tackle classification while training faster than BP based ANNs consisting one
and regression tasks by implementing them within other hidden layer on large datasets thanks to its randomization
networks. The paper reviewed in this study that employs mechanism. ELM networks have been criticized for being
RBMs uses it in a setting where RBMs are followed by a unoriginal (Wang & Wan ).

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
11 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Figure 8 | RBM and DBN examples.

DEEP Q NETWORKS (DQN) (Figure 9). We explain this aspect with their respective suc-
cess in matrix prediction and sequence prediction, tasks that
DQN (Mnih et al. ) is a reinforcement learning algor- have high importance in hydrologic modelling. One con-
ithm in which there is a predetermined rewarding policy. founding thing is that even though LSTM networks were
The DQN works similarly to the Q-Learning algorithm vastly employed, one architecture that yields similar per-
which works by providing a medium for the agent and the formance, GRU networks did not find significant usage in
environment and tries to maximize the rewards the agent the field. Additionally, we found it surprising that even
gets through its actions in the environment. A DQN differen- though most of the studies reviewed here tackles tasks invol-
tiates from the conventional Q-Learning algorithm with how ving sequential data, Transformers were not employed by
it generalizes. In other words, the agent in the Q-Learning any of the studies we reviewed. It should be noted that a
algorithm cannot estimate the outcome from action-environ-
ment pairs it didn’t see before, while in DQN the agent is
able to produce a reinforced action.

SUMMARY OF ARTICLES

This subsection presents visual summaries of the reviewed

papers as well as a summary table extracted information
from each paper that received comprehensive review. As
previously stated, information from a total of 129 papers
are given, however data points in shared figures do not
always equal to this value depending on the data column,
this is because some entries in the table may have multiple
values in some data columns instead of one.
Among the reviewed papers, as in the deep learning lit-
erature, most used architectures are CNNs and LSTMs Figure 9 | Architectures described in the study and their usage in reviewed papers.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
12 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Transformer is a neural network architecture that is vastly (Figure 14). Figure 15 summarizes the usage of numerical
used in the field of natural language processing, which is computing frameworks. Even though TensorFlow seems to
another field that focuses on sequential data. be the first choice among water domain researchers, it
As the utilization of deep learning in the water field in a should be noted that most of the usage comes from Keras,
broad sense increases over time (Figure 10), annual usage of the second most used framework, which typically works
the deep neural network architectures increases. Figure 11 on top of TensorFlow by providing a higher-level interface.
shows the change in usage through March 2020 and also Thus, our inference is that Keras is the most used deep learn-
presents simple projections for the rest of the year built on ing framework within the water field. In contrast, libraries
top of the number of publications through March 2020. like PyTorch that is highly endorsed in deep learning litera-
We expect to see growth in the usage of neural network ture find a smaller place to themselves. Reason of this might
architectures that have been widely used in other disciplines be Keras’ ability in fast ANN prototyping and modeling with
but not in the water field like DQNs and GANs. least implementation-relevant details possible.
The datasets used in publications reviewed typically are
datasets acquired from authorities or governmental agencies
(Figure 12). Even though in deep learning literature, the RESULTS
dataset acquisition primarily done by using previously exist-
ing datasets, in the water field, on the contrary, this does not In this section we provide brief summaries of papers pre-
seem to be the case. Code accessibility of the papers is sented in the previous section (Table 1.) Papers are
another aspect of the studies published in the water field grouped and interpreted by their use case.
that differentiates it from the deep learning field in general.
Although open-sourced models are widely expected from
deep learning researchers, open-sourcing the software built STREAMFLOW AND FLOOD
for a study is unusual for a publication in the water field if
not rare, as it can be seen in Table 1. Cumulating from Runoff prediction and flood forecasting are major tasks in
both the data acquisition type and code accessibility, repro- rainfall-runoff modeling. Toward this end, many researchers
ducing the outcomes of a papers does not seem to be an easy have applied cutting-edge deep learning architectures to the
task for the authorities and other researchers in the field. runoff prediction and flood forecasting tasks. Since rainfall
Owing to the fact that the field of hydrology vastly relies and runoff are both time series data, the common net-
on sequential data, most studies seem to work on sequence works for the streamflow prediction and flood forecast
prediction and regression tasks (Figure 13). This phenom- are RNN, LSTM, NAR, and ENN. Kratzert et al. ()
enon also might be linked with the fact that most of the applied an LSTM model on daily runoff prediction for the
studies reviewed were classified in Flood subdomain first time, considering the meteorological observations
with results better than a well-established physical model
SAC-SMA þ Snow-17. In 2019, Kratzert et al. (2019) further
applied the LSTM model to 531 watersheds in U.S. with
k-fold cross validation and it shows that LSTM can be applied
on ungaged watersheds with better results than physical
models such as calibrated SAC-SMA and the National
Water Model. Other researchers applied recurrent neural net-
works to the runoff forecast and compared their outputs to
other machine learning models. Damavandi et al. () pro-
posed an LSTM model on a Texas watershed predicting the
next day’s daily streamflow using climate data and the current
day’s streamflow. Their results show that LSTM performs
better than physical model CaMa-Flood. Zhang et al.
(e) applied an LSTM model on monthly reservoir
inflow and outflow predictions in hourly, daily, and monthly
Figure 10 | Average number of papers published each month during 2018, 2019 and
basis with better results than SVM and BPNN. Kumar et al.
2020 January to March. () applied RNN and LSTM models for the monthly

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
13 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Figure 11 | Architectures described and their annual usage in reviewed papers.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
14 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

(SVD). Zuo et al. () proposed three LSTM models

with different pretreatments, which are variational mode
decomposition (VMD), ensemble empirical mode decompo-
sition (EEMD), and discrete wavelet transform (DWT) for
daily streamflow up to 7 days ahead. Zhu et al. (2020) pro-
posed a probabilistic LSTM model coupled with the
Gaussian process (GP) to deal with the probabilistic daily
streamflow forecasting. These studies show that the pretreat-
Figure 12 | Distribution of dataset acquisition types in reviewed papers. ment of input may help to improve the deep learning model
accuracy.
Other studies focused on constructing more complex
rainfall prediction in India and shows LSTM provides better deep learning architectures such as autoencoder, enco-
results. Qin et al. () applied the LSTM on the streamflow der-decoder, and customized layer based on the LSTM.
prediction and compared it with the Autoregression model. Bai et al. () proposed an LSTM model with stack auto-
Yang et al. (a) applied an LSTM model on the daily reser- encoder (SAE) to predict daily discharge values based on
voir overflow prediction. NAR with external input and LSTM one-week discharge values, and the results of SAE-LSTM
are used for three reservoirs with geographic information, outperform LSTM alone. For multiple time-step flood fore-
daily precipitation, air temperature, wind speeds, relative cast tasks, an encoder-decoder LSTM is proposed for the
humidity, and sunshine duration. Wan et al. () also suc- runoff prediction by Kao et al. () and Xiang et al.
cessfully applied Elman Neural Network to a real-time 3- (). Kao et al. () proposed an encoder-decoder
hour ahead flood forecasting. LSTM model that can be used on multi-timestep output pre-
Since hyper-parameter optimization is a problem in dictions for up to 6 hours. Xiang et al. () proposed an
deep learning, some studies focused on applying encoder-decoder LSTM model that can be used to predict
additional optimization algorithms on deep learning for up to 24 hours ahead. Both studies showed the enco-
models. Yuan et al. () proposed two models that use der-decoder LSTM is better than LSTM. In particular,
the ant lion optimizer (ALO) and particle swarm optimiz- Kratzert et al. (2019) proposed the Entity-Aware-LSTM net-
ation (PSO) to optimize the parameters and hidden layers work, which designed a specific layer for the rainfall-runoff
of an LSTM model, which are named LSTM-ALO and modeling based on LSTM. This network allows for learning
LSTM-PSO. Their results show that the LSTM-PSO outper- catchment similarities as a feature layer, and data from mul-
formed the LSTM and LSTM-ALO. Yang et al. (a) tiple watersheds can be used to provide runoff for a
proposed a Genetic Algorithm based NAXR (GA-NAXR), watershed. These studies applied some high-level designs
which outperforms the NAXR and LSTM. Ni et al. () to the LSTM network and can perform much better than
proposed two LSTM based models, which are wavelet- normal LSTM on the long-term or multiple watersheds.
LSTM (WLSTM) and CNN þ LSTM (CLSTM) for the rain- Several studies coupled physical models with deep
fall and streamflow forecasting, and both models have learning networks.Yang et al. (b) proposed a model
better results than LSTM. Kabir et al. () proposed a using LSTM to improve the performance of flood simu-
wavelet-ANN to make the hourly streamflow predictions. lations of a physical model. The watershed-averaged daily
Results show that wavelet-ANN can make acceptable pre- precipitation, wind, temperature and model-simulated dis-
dictions for at most 6 hours ahead, which outperforms charge from GHMs þ CaMa Flood model in 1971–2020
ANN, DBN, and SVR. were used to model the discharge. This indicates the deep
Some studies focused on the data pretreatment such learning models can be used to improve the streamflow
as the decomposition of rainfall and runoff prior to deep forecast accuracy of physical models. Nguyen & Bae
learning models. He et al. (2019) proposed a DNN model () proposed an LSTM network using the quantitative
for daily runoff prediction where the inputs are the runoff precipitation forecasts (QPFs) of the McGill Algorithm for
series that were decomposed into multiple intrinsic mode Precipitation nowcasting by Lagrangian Extrapolation
functions (IMFs) with variational mode decomposition (MAPLE) system to reproduce three-hour mean areal pre-
(VMD). Hu et al. () proposed an LSTM model for cipitation (MAP) forecasts. Corrected MAPs are used as
flood forecast with the preprocessing of proper orthogonal input to a coupled 1D/2D urban inundation model to pre-
decomposition (POD) and singular value decomposition dict water levels and relevant inundation areas (1D

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Table 1 | Reviewed papers with curated data points

15
M. Sit et al.
Paper Network Type Framework Dataset Open Source Reproducible DL Task Water Field

Yuan et al. () LSTM MATLAB Acquired No No Sequence Prediction Flood

Zhang et al. (e) LSTM – Acquired No No Sequence Prediction Flood

|
A comprehensive review of deep learning applications in hydrology
Kratzert et al. () LSTM Keras, TensorFlow Existing No Yes Sequence Prediction Flood
He et al. (b) ANN, DBN – Acquired No Yes Sequence Prediction Flood
Hu et al. () LSTM – Acquired No No Regression Flood
Wang et al. (b) CNN – Acquired No No Sequence Prediction Flood
Yang et al. (a) NAR, LSTM Keras, TensorFlow Acquired No No Regression Flood
Sankaranarayanan et al. () ANN Keras Acquired No No Classification Flood
Ni et al. () LSTM, CNN – Acquired No No Sequence Prediction Flood, Weather
Bai et al. () LSTM, AE – Acquired No No Sequence Prediction Flood

Uncorrected Proof
Yang et al. (b) LSTM – Acquired No No Regression Flood
Bhola et al. () CNN – Collected No No Segmentation Flood
Damavandi et al. () LSTM Keras, TensorFlow Acquired No Yes Sequence Prediction Flood
Moy de Vitry et al. () CNN TensorFlow Collected Yes Yes Matrix Prediction Flood
Worland et al. () ANN Keras, TensorFlow Acquired No No Regression Flood
Kratzert et al. (a) LSTM PyTorch Acquired Yes Yes Sequence Prediction Flood
Kumar et al. () RNN, LSTM Keras Acquired No Yes Sequence Prediction Flood
Wan et al. () ENN – Acquired No No Sequence Prediction Flood
Qin et al. () LSTM TensorFlow Acquired No No Sequence Prediction Flood
Kratzert et al. (b) LSTM PyTorch Existing Yes Yes Regression Flood
Bui et al. (a) ANN MATLAB Acquired No No Classification Flood
Nguyen & Bae () LSTM TensorFlow Acquired No No Sequence Prediction Flood
Bui et al. (b) ANN – Acquired No No Classification Flood
Kabir et al. () DBN TensorFlow Acquired No No Sequence Prediction Flood

Water Science & Technology

Kao et al. () LSTM Keras Acquired No No Sequence Prediction Flood
Xiang et al. () LSTM Keras, TensorFlow Acquired No No Sequence Prediction Flood
Zuo et al. () LSTM Matlab, TensorFlow Acquired Yes Yes Sequence Prediction Flood
Ren et al. () ANN, LSTM, GRU – Acquired No No Sequence Prediction Flood
Zhu et al. (b) LSTM Keras Acquired No Yes Sequence Prediction Flood
Wang et al. (d) CNN Keras Acquired No No Classification Flood
Laloy et al. () GAN Lasagne, Theano Existing No No Matrix Prediction Groundwater

|
in press
Wang et al. (b) ANN – Collected Yes No Matrix Prediction Groundwater
Santos et al. () CNN Keras Existing Yes Yes Matrix Prediction Groundwater

|
2020
(continued)

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

 Table 1 | continued

16
Paper Network Type Framework Dataset Open Source Reproducible DL Task Water Field

M. Sit et al.
Mo et al. (a) CNN PyTorch Collected Yes Yes Matrix Prediction Groundwater
Mo et al. (b) CNN PyTorch Collected Yes Yes Matrix Prediction Groundwater

|
Sun et al. (a) CNN Keras, TensorFlow Acquired No No Sequence Prediction Groundwater

A comprehensive review of deep learning applications in hydrology

Jeong & Park () NARX, LSTM, GRU TensorFlow Acquired No No Sequence Prediction Groundwater
Zhou et al. () CNN PyTorch Acquired No No Regression Groundwater
Jeong et al. () LSTM TensorFlow Acquired No No Sequence Prediction Groundwater
Zhang et al. (d) CNN Keras, TensorFlow Collected No No Segmentation Land Use and Soil
Zhang et al. ( g) LSTM TensorFlow Acquired No No Regression Land Use and Soil
Cao et al. () CNN, ELM – Existing No No Segmentation Land Use and Soil
Zeng et al. () ANN – Acquired No No Regression Land Use and Soil
Reddy & Prasad () LSTM – Acquired No Yes Sequence Prediction Land Use and Soil

Uncorrected Proof
Fu et al. () CNN – Acquired No No Classification Land Use and Soil
Jiang () AE – Acquired No No Classification Land Use and Soil
Shen et al. () ANN R/H2O Acquired No No Regression Land Use and Soil
Jin et al. () CNN TensorFlow Acquired No No Segmentation Land Use and Soil
Persello et al. () CNN – Acquired No No Segmentation Land Use and Soil
Kroupi et al. () CNN TensorFlow Existing No No Classification Land Use and Soil
Sun et al. (b) CNN PyTorch Acquired No No Classification Land Use and Soil
Meng et al. () CNN – Acquired No No Classification Land Use and Soil
Kopp et al. () CNN – Collected No No Segmentation Land Use and Soil
Jiang et al. () CNN – Acquired No Yes Matrix Prediction Land Use and Soil
Bhosle & Musande () CNN TensorFlow Existing No No Classification Land Use and Soil
Zhang et al. (a) ANN, CNN – Existing No No Matrix Prediction Land Use and Soil
Wang et al. (a) DBN – Acquired No No Matrix Prediction, Regression Land Use and Soil
Nam & Wang () AE R/H2O Acquired No No Classification Land Use and Soil

Water Science & Technology

O’Neil et al. () CNN TensorFlow Acquired Yes Yes Classification Land Use and Soil
Yang et al. () CNN Keras Acquired No Yes Segmentation Land Use and Soil
Abdi et al. () CNN – Existing No Yes Classification Others
Li et al. () CNN, AE PyTorch Acquired No Yes Segmentation Others
Rohmat et al. () ANN MATLAB Acquired No No Regression Others
Amirkolaee & Arefi () CNN MATLAB Existing No Yes Matrix Prediction Others

|
Wang et al. (a) CNN Keras Existing No No Classification Others

in press
Kylili et al. () CNN – Collected No No Classification Others

|
Kang () CNN Keras, TensorFlow Acquired No No Classification Others

2020
Haklidir & Haklidir () ANN – Acquired No Yes Regression Others

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

 Kim et al. () GAN, AE – Acquired No No Unsupervised Learning Others

17
Zhang et al. (f) LSTM Theano Acquired No No Regression Surface Water

M. Sit et al.
Liu et al. (b) ANN Keras, TensorFlow Acquired No Yes Sequence Prediction Surface Water
Xiao et al. (a) CNN, LSTM Keras, TensorFlow Acquired No No Matrix Prediction Surface Water
Read et al. () LSTM – Acquired No No Regression Surface Water

|
A comprehensive review of deep learning applications in hydrology
Liu et al. (a) CNN Caffe Existing No No Segmentation Surface Water
Xiao et al. (b) CNN PyTorch Existing No No Classification Surface Water
Mei et al. () CNN PyTorch Existing No No Regression Surface Water
Ling et al. () CNN MATLAB Acquired No No Segmentation Surface Water
Song et al. () CNN Caffe Existing No No Classification Surface Water
Hrnjica & Bonacci () LSTM – Acquired No No Sequence Prediction Surface Water
Qi et al. () LSTM – Acquired No No Sequence Prediction Surface Water
Zhu et al. (a) LSTM ANNdotNET, CNTK Acquired No No Sequence Prediction Surface Water

Uncorrected Proof
Lee & Lee () RNN, LSTM – Acquired No No Regression Water Quality
Hamshaw et al. () RBM MATLAB Acquired Yes Yes Classification Water Quality
Liu et al. () LSTM Keras, TensorFlow Acquired No No Sequence Prediction Water Quality
Yurtsever & Yurtsever () CNN Caffe Collected No No Classification Water Quality
Li et al. () LSTM, GRU, ENN – Acquired No No Sequence Prediction Water Quality
Shin et al. () LSTM – Acquired No No Sequence Prediction Water Quality
Banerjee et al. () ANN R/H2O Collected No No Regression Water Quality
Wang et al. (c) LSTM – Acquired No No Sequence Prediction Water Quality
Yim et al. () ANN MATLAB Collected No No Regression Water Quality
Zou et al. () LSTM Keras Acquired No No Sequence Prediction Water Quality
Liang et al. () LSTM Keras Acquired No No Sequence Prediction Water Quality
Barzegar et al. () CNN, LSTM – Collected No No Sequence Prediction Water Quality
Yu et al. () LSTM – Acquired No No Sequence Prediction Water Quality
Zhang et al. (a) LSTM Keras, TensorFlow Acquired No No Sequence Prediction Water Resources Management

Water Science & Technology

Zhang et al. (b) LSTM, GRU Keras, TensorFlow Acquired No No Sequence Prediction Water Resources Management
Zhang et al. (c) LSTM, NARX, ENN Keras, TensorFlow Collected Yes Yes Regression Water Resources Management
Harrou et al. () DBN – Collected No No Sequence Prediction Water Resources Management
Shi & Xu () AE – Collected No No Regression Water Resources Management
Zhou et al. () ANN, CNN PyTorch Existing Yes Yes Classification Water Resources Management
Fang et al. () CNN – Acquired No No Classification Water Resources Management
Sun et al. () ANN – Acquired No No Regression Water Resources Management

|
in press
Karimi et al. () ANN, LSTM MATLAB Collected No No Sequence Prediction Water Resources Management
Xu et al. () LSTM – Acquired No No Regression Water Resources Management

|
2020
(continued)

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

 18
M. Sit et al.
|
|

A comprehensive review of deep learning applications in hydrology

Table 1 continued

Paper Network Type Framework Dataset Open Source Reproducible DL Task Water Field

Nam et al. () DQN – Acquired No No Reinforcement Learning Water Resources Management
Mamandipoor et al. () LSTM Keras, TensorFlow Acquired No No Classification Water Resources Management
Tang et al. () ANN – Acquired No No Regression Weather
Klampanos et al. () AE, CNN – Acquired No No Unsupervised Learning Weather
Scher & Messori () CNN Keras, TensorFlow Acquired No No Matrix Prediction Weather
Ukkonen & Mäkelä () ANN Keras Acquired No Yes Classification Weather

Uncorrected Proof
He et al. (a) LSTM – Acquired No No Classification Weather
Jeppesen et al. () CNN Keras, TensorFlow Existing Yes Yes Segmentation Weather
Chen et al. () GAN PyTorch Acquired No No Matrix Prediction Weather
Wieland et al. () CNN Keras, TensorFlow Existing No Yes Segmentation Weather
Weyn et al. () CNN, LSTM Keras, TensorFlow Acquired No No Matrix Prediction Weather
Wei and Cheng, (2019) ANN Weka Acquired No No Sequence Prediction Weather
Zhang et al. (b) CNN – Acquired No Yes Classification Weather
Kim et al. () ANN – Acquired No No Regression Weather
Pan et al. () CNN – Acquired No Yes Matrix Prediction Weather
Tran & Song () LSTM, GRU, CNN TensorFlow Existing No Yes Matrix Prediction Weather
Poornima & Pushpalatha () LSTM, ELM Keras, TensorFlow Acquired No Yes Sequence Prediction Weather
Chai et al. () CNN – Acquired No Yes Segmentation Weather
Wu et al. () LSTM, CNN – Acquired No Yes Matrix Prediction Weather
Zhang et al. () LSTM – Acquired No No Regression Weather

Water Science & Technology

Su et al. () LSTM, CNN – Acquired No No Sequence Prediction Weather
Chen et al. () CNN, LSTM PyTorch Existing No Yes Matrix Prediction Weather
Weber et al. () CNN TensorFlow Acquired Yes Yes Matrix Prediction Weather
Yan et al. () CNN TensorFlow Acquired No Yes Sequence Prediction Weather
Wang et al. (c) RNN – Acquired No No Matrix Prediction Weather

|
in press
|
2020
Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf
 Uncorrected Proof
19 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

conduit network model and 2D overland flow model). In

this model, LSTM takes a forecast from MAPLE and then
reproduced MAPs are used by coupled 1D/2D urban inun-
dation model. Worland et al. () proposed a DNN
model to predict the flow-duration curves using USGS
streamflow data by constructing 15 output values represent-
ing 15 quantiles of the curve. These indicate the deep
learning models can be used as surrogate models of phys-
ical models or curves.
Some studies focused on flood susceptibility and flood
prediction capabilities. Bui et al. (2020) applied a DNN net-
work with 11 variables such as DEM, aspect, slope, etc. to
predict the flood susceptibility, a value between 0 and 1,
for an area. Wang et al. (2020) proposed a CNN network
Figure 13 | Histogram of machine learning task types studied in reviewed papers. with 13 flood triggering factors in the study area to map
the flood susceptibility for the study area. In this project,
each pixel may have different flood susceptibility values.
Sankaranarayanan et al. () applied a DNN network to
classify if flooding is possible in an area for the month
given total precipitation and temperature. This study simpli-
fied the project into a binary prediction task of if there will
be a flood event or not. The same simplification is done by
Tien et al. (2020), who applied a DNN network to classify
if a region is flash flood prone or not using the elevation,
slope, aspect, curvature, stream density, NDVI, soil type,
lithology and rainfall data as input.
With CNNs, some innovative flood monitor and fore-
cast projects are now possible. Wang et al. (2019) applied
a CNN model that predicted the real-time hourly water
levels for warning systems during typhoons using satellite
images. Bhola et al. () proposed a model that estimates
Figure 14 | Histogram of Water domain subfields studied in reviewed papers. the water level from CCTV camera images using a CNN
model. The model performed the edge detection to find the
water body in the photo, and the water level is then calcu-
lated from some physical measurements. Moy et al. (2019)
proposed a model that calculated the flooded-area and the
surface observed flooding index with CCTV cameras using
CNN. U-net CNN is used to segment the stream shape
from a CCTV camera and another CNN model is used to cal-
culate the flooded areas in pixels from the segmented photo.

SUBSURFACE AND GROUNDWATER

There are several different types of studies in subsurface and

groundwater. Different deep learning models can be used in
each of different types of studies. One type of groundwater
study is the estimation of water table level or flow rate.
Figure 15 | ML frameworks described in the study and their usage in reviewed papers. With the groundwater monitoring wells data, this is a 1-D

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
20 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

regression task similar to the surface water predictions using contaminate source in groundwater. Some complex tra-
recurrent models. Jeong & Park () applied the NARX, ditional algorithms can be used to solve this problem.
LSTM and GRU on the water table level estimations with However, it will be not efficient when the data is in high
observed data from monitoring wells. It is found that the dimensions. Mo et al. (b) developed a deep autoregres-
estimations from the NARX and LSTM models are superior sive neural network which was used as the surrogate
to those of the other models in terms of prediction accuracy. model of this high dimensional inverse problem and pro-
Jeong et al. () further applied LSTM with multiple loss vides more accurate inversion results and predictive
functions, which shows that the proposed LSTM model uncertainty estimations than physical models. Laloy et al.
with cost function of MSE with Whittaker smoother, least () proposed another approach for the inversion by
trimmed squares, and asymmetric weighting has the best using GAN. With the GAN models, the inversion rapidly
performance on groundwater level prediction with data explores the posterior model distribution of the 2-D steady
corrupted by outliers and noise. For groundwater flow pre- state flow and recovers model realizations that fit the data
diction, Wang et al. (2020) proposed a theory-guided deep close to the true model results.
neural network which not only makes the predictions of Finally, Sun et al. () investigate three learning-based
groundwater flow, but also estimates the parameters of the models, namely DNN, Multiple Linear Regression (MLR),
partial differential equation (PDE) as well as the initial con- and SARIMAX, to find missing monthly data in total
dition and boundary condition of the PDE. By constructing water from GRACE (Gravity Recovery and Climate Exper-
the loss function, the theory-guided neural network can pro- iment) Data (Tapley et al. ). Based on the results, the
vide the prediction results with reasonable parameters of the performance of DNN is slightly better than SARIMAX, sig-
physical model PDE. nificantly better than MLP in most of the basins. However,
More groundwater studies take a cross section in time- three learning-based models are reliable for the reconstruc-
series and focus on a 2-D map. Examples include the tion of GRACE data in areas with humid and no/low
groundwater water balance map (Sun et al. 2019), hydraulic human interventions. Dynamic multiphase flow in hetero-
conductivity field map (Zhou et al. ), pressure and CO2 geneous media is a hard problem in groundwater studies.
saturation field map (Mo et al. a). CNN models can be These studies show the deep learning models can work as
used in these studies. Sun et al. (2019) proposed a CNN surrogate models by improving computational efficiency in
model using the physical model NOAH simulation results groundwater dynamic predictions and inversions.
as input to predict the groundwater water balance. Thus,
the CNN model is used to correct the physical model results
by learning the spatial and temporal patterns of residuals LAND AND SOIL
between GRACE observations and NOAH simulations.
Results show that the CNN can significantly improve the The segmentation and classification of land use and land
physical model simulation accuracy. Zhou et al. () pro- cover are important for water and soil resources manage-
posed a CNN of eight layers to learn a map between ment. Many studies have applied the deep learning
stochastic conductivity field and longitudinal macro- networks to create a more accurate land cover map from
dispersivity based on synthetic 2-D conductivity fields. The satellite or radar imagery. Abdi et al. () applied SAE
estimations are in acceptable accuracy with moderate and CNN on land cover classification of urban remote sen-
heterogeneity. Mo et al. (a) proposed an Encoder- sing data. Tests on 9 datasets show that SAE and CNN are
Decoder CNN to approximate the pressure and CO2 satur- better than machine learning models like logistic regression,
ation field map in different time steps as a surrogate model. Naive Bayes, KNN and SVM. Cao et al. () applied ELM
Furthermore, some studies in groundwater concern 3D and CNNs to classify land cover categories on satellite ima-
mapping. An example is the flow rate estimation in a 3D gery and the results show the combined CNN–ELM method
rock. Santos et al. () proposed a 3D CNN to predict has the highest accuracy. Shen et al. () applied a DNN
the estimated state solution of Navier-Stokes equation for network to predict the land drought index using the precipi-
laminar flow, which is the flow rate, with 3D rock images. tation data, meteorological drought index data and soil
This 3D CNN is a surrogate model of the Navier-Stokes relative moisture data. Bhosle et al. (2019) applied a CNN
equation, and takes less than a second. model on the land cover land use classification on Indian
In groundwater studies, one of the other deep learning Pines dataset. Some studies applied more complex coupled
applications is inversion, such as to identify the models including encoder-decoder, autoencoder, 3D

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
21 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

networks, and coupled machine learning models to achieve two CNN models, SegNet (Badrinarayanan et al. ) and
a higher model accuracy. Zhang et al. (2019) applied 3D- VGG16 (Simonyan & Zisserman ), for the segmentation
CNN and 3D-DenseNet models on the land cover land and classification of agricultural fields. Sun et al. (2019)
use classification on Indian Pines and Pavia University data- applied a CNN model to classify if the land is impervious
set. All these studies showed a high accuracy of the land surfaces, vegetation or bare soil from satellite imagery or
cover identification task. Kounghoon et al. (2020) applied both the satellite and LiDAR data. Results show that the
the autoencoders prior to the random forest, SVM and model with LiDAR data provides better results than the
other machine learning models to the landslide suscepti- model with satellite imagery only. Meng et al. () applied
bility prediction. Results show that random forest with an to identify the wetland types in one lake using the Chinese
autoencoder gives more accurate results than other machine remote sensing imagery GF-2. Bhosle et al. (2019) applied
learning methods. In the study of wetland type identification a CNN model to identify crop types on EO-1 Hyperion
by Meng et al. (), results show that the ensemble model sensor hyperspectral imagery. Yang et al. () applied
SVM-CNN performs better than CNN and SVM. O’Neil CNNs to identify the water body types from remote sensing
et al. () applied an Encoder-Decoder CNN model to images. Mask R-CNN is used to segment the water body,
classify the wetland types using LiDAR radar dataset and and the ResNets (He et al. ) including ResNet-50 and
NDVI index data. This study used the LiDAR DEM and ResNet-101 are used to identify the water body types. Results
remote sensing images to generate physically informed show a high accuracy on regular-shaped water bodies.
input including Slope, NDVI, DTW, and TWI. In particular, Similar networks can be used to generate DEMs,
Kroupi et al. () applied a CNN based model to land which is another type of land use study. Jiang et al. ()
cover classification that trained on the European satellite applied a CNN model to predict the paleo-valley DEM
dataset – EuroSAT and tested in a region outside Europe, using the original DEM data and electrical conductivity.
and still provides promising results despite differences in The electrical conductivity data are collected from the field
tested and trained regions. These results indicate the CNN study, and the CNN models used in this study distinguish
models have a high model accuracy as well as high robust- the valley and non-valley pixels, which can find the spatial
ness in the land use land type classification task. connectivity of the paleo-valley. The CNN can efficiently con-
In addition to applying some known networks, some strain three-dimensional paleo-valley geometry DEM.
researchers developed objective-based CNN rather than Snow cover, a special study in land cover studies, can
traditional pixel-based CNN to better identify the land be measured by deep learning models in different
use and land types. Zhang et al. (2018) proposed an approaches. Kopp et al. () proposed a model to predict
object-based CNN to label very fine spatial resolution the snow depth using OpenCV and Mask R-CNN on the sur-
(VFSR) remotely sensed images to do object-wise segmenta- veillance camera photos. The Mask R-CNN is used to
tion rather than pixel-wise segmentation in urban areas. Fu segment the detectable measuring rod, and then the
et al. () proposed a blocks-based object-based CNN for OpenCV library can be used to identify the snow depth by
classification of land use and land cover types and achieved measuring how much of the measuring rod gets covered by
end-to-end classification. This model works well on irregular snow. Wang et al. (2020) applied a DBN and CNN to estimate
segmentation objects which is a common in land use classi- snow depth from the geographical data. DBN takes multiple
fication. Jiang () proposed an object-based CNN with an inputs including latitude, longitude, elevation, forest cover
autoencoder for extracting high level features automatically. fraction, time, passive microwave horizontal and vertical
Results show the AE-object-CNN is better than three polarization brightness temperatures of 19 and 37 GHz.
manual design feature systems. Jin et al. () proposed Results show that the DBN outperforms CNN in this study.
an object-oriented CNN which used a typical rule set of fea- Additional land and soil related studies include the
ture objects to construct the object-oriented segmentation land surface temperature, soil salinity, vegetation dynamics
results before using a CNN to make the classification. over time, and these can be done with time-series related
These studies show object-based studies have better per- models such as LSTM. Zhang et al. (2018) applied the
formance than simple CNN. ensemble composition-based LSTM models on the daily
Some studies focus on more specific tasks based on land surface temperature simulation. The original daily land
different study purposes such as the identification of agri- surface temperature data series were decomposed into
cultural fields, impervious surfaces, wetland types, water many Intrinsic Mode Functions (IMFs) and a single residue
body types, and crop types. Persello et al. () applied item. And the Partial Autocorrelation Function (PACF) is

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
22 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

used to obtain the number of input data sample points for from the LSTM models with different decomposed inflow
LSTM models. Zeng et al. () applied Partial Least data as inputs for more accurate assumptions. Zhang et al.
Square Regression (PLSR), SVM and DNN for predicting (2018) propose an LSTM model to predict water table
soil salinity from images. Surprisingly, DNN performs depth in agricultural areas. The LSTM model takes monthly
worse than PLSR for this task. Reddy & Prasad () applied water diversion, evaporation, precipitation, temperature,
a LSTM model to predict the vegetation dynamics using and time as inputs for prediction of the water table depth.
NDVI images from 2000 to 2016 in a 7-day gap. This study The results of the LSTM model are compared with the
shows that the single feature NDVI can be used to provide FFNN model, and the LSTM model performs better than
accurate vegetation prediction over time. the FFNN model. Also, it is highlighted that the dropout
method increases the LSTM model’s accuracy in this task.
Alongside the water level or flow, the prediction of
SURFACE WATER water temperature has received much attention in the
scholarship. Xiao et al. (a) propose a convolutional
The prediction of water level is crucial for water resources long short-term memory (ConvLSTM) model to predict the
management and protecting the natural environment. sea surface temperature (SST). In the paper, 36 years of sat-
Many studies have applied deep learning methods, such ellite-derived SST data are used. Based on the results, the
as LSTM and ANN, to forecast the water level from one ConvLSTM model produces more accurate results than the
day to one year. Hrnjica & Bonacci () investigate two linear support vector regression model and two different
different ANNs, namely LSTM and Feed Forward Neural LSTM models for short and mid-term temperature prediction.
Network (FFNN), to forecast the lake water level. Monthly Read et al. () aim to predict lake water temperatures based
measurements for the last 38 years of Vrana Lake, Croatia on depth. A hybrid LSTM and theory-based feedbacks (model
is utilized for training the models in order to predict 6 or penalties) model was developed. According to the results, the
12 months ahead of the lake water level. The set of hybrid model produces the best results among the tested
sequences with different lengths created from the obtained methods in the paper. The results from the paper can be
data is used in the networks, instead of using classical seen as an example of improving predictions by integrating
lagged data. The results of LSTM and FFNN are compared scientific rules with DL methods.
with classical time forecasting methods and ANN. Accord- In addition to the aforementioned subtopics in surface
ing to the results, the performance of LSTM is the best water, various tasks are also investigated by scholars such
among the models in all scenarios, while FFNN provides as segmentation, change detection, or super-resolution.
better accuracy than the compared methods in both 6 and Despite extensive usage of LSTM or ANN in previously
12 months prediction. Zhu et al. (a) also investigate aforementioned papers, CNN models are generally used in
LSTM and FFNN in their work with data from a different these tasks. Liu et al. (b) label objects in aerial images
region. However, the models in this paper are designed to pre- as water, tree, etc. with the help of a fully convolutional
dict one month ahead of the lake water level for 69 temperate neural network (FCN) model and multinomial logistic
lakes in Poland. Their results indicate that LSTM and FFNN regression. FCN takes the aerial image and returns a prob-
perform similarly most of the time, unlike in the previous ability. At the same time, LiDAR data passes into
paper. The reasons for this situation can be the differences multinomial logistic regression and returns another prob-
between the datasets, prediction intervals, or model designs. ability. These two probabilities are combined with higher-
In addition to lake water level prediction, Liu et al. order conditional random field formulation to produce a
(b) developed a Bayesian Neural Network, which is final probability. Based on the final probability, objects are
based on ANN with posterior inference, to forecast the labeled with the corresponding group. Mei et al. ()
water level in a reservoir to derive operation rules. According develop a CNN based model to measure the sea ice thick-
to the paper, the current reservoir status and future inflows ness in Antarctica from LiDAR data. The input is a
are the primary factors that affect operational decisions. windowed LiDAR scan (snow freeboard), and the mean
Also, the influence of inflow uncertainties on reservoir oper- ice thickness is the output of the model. In addition to the
ations is more than model parameter uncertainty. Their LiDAR scan, the paper investigates the effects of different
findings show the impact of the input data alongside the inputs such as surface roughness and snow depth on the
promising results of Bayesian NN. Qi et al. () forecast task. Ling et al. () use the CNN model in order to gener-
daily reservoir inflow by ensembling the different results ate a finer resolution of the image to measure the wetted

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
23 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

river width. Then, the output of CNN is used to measure the as pH and DO. Liang et al. () also work with the predic-
width of the river. Song et al. () aim to detect the change tion of the chlorophyll-a concentration level. Fabricated
in surface water in remote sensing images. An FCN model is data are created by the environmental fluid dynamics code
proposed and used for the same regions at different times. (EFDC) to train an LSTM model. The LSTM model can
The FCN model returns the surface water regions for each forecast the chlorophyll-a concentration level for up to one
time. The outputs for the same regions are compared, and month. Chlorophyll-a, water temperature, and total phos-
a change map is created to show the difference. Beyond phorus are identified as critical inputs that affect the
these publications, Xiao et al. (b) introduce a dataset performance of the LSTM model. Barzegar et al. () inves-
for the classification of ice crystals. The dataset contains tigate multiple models to predict the level of DO and
7,282 images in 10 different categories. The performance of chlorophyll-a in Small Prespa Lake, Greece. Three different
numerous pre-trained models, such as AlexNet (Krizhevsky NN models, namely CNN, LSTM, and CNN-LSTM, were
et al. ) and VGGNet(s) (Simonyan & Zisserman ), developed to forecast the DO and chlorophyll-a concen-
are also provided in the paper. trations and use pH, oxidation-reduction potential (ORP),
water temperature, and electrical conductivity (EC) as inputs
for the models. In addition to ANN models, SVM and decision
WATER QUALITY tree models are used for the performance comparison. Accord-
ing to results, the hybrid CNN-LSTM model provides the best
Water quality monitoring and prediction are vital oper- accuracy to predict both DO and chlorophyll-a.
ations for many fields, such as water resources Yim et al. () develop a stacked autoencoder-deep
management and water treatment. Water quality and safety neural network (SAE-DNN) to predict phycocyanin (PC)
depend on numerous parameters with complex biological, concentrations in inland waters from in-situ hyperspectral
chemical, and physical interactions. As such, deterministic data. The proposed architecture’s ability for the prediction
water quality models are a realistic option in only the simplest from airborne hyperspectral imagery is examined. Shin
and idealized scenarios. However, data-driven models are et al. () introduce an LSTM model to forecast the occur-
increasingly being used in a variety of water quality appli- rence of harmful algal blooms in the South Sea of Korea.
cations. One such application area is predicting surface Sea surface temperature and photosynthetically available
water quality. Li et al. () propose an ensemble approach radiation are extracted from satellite data to be used as
that combines three RNN models with Dempster/Shafer (D- inputs for the LSTM model in order to minimize the
S) evidence theory (Shafer ) to predict the quality of damage. Lee & Lee () aim to predict the occurrence
water. The results of three RNN models, namely LSTM, and number of harmful algal blooms with an LSTM model
GRU, and Elman Neural Network, are combined by D-S evi- by providing weekly water quality and quantity data. The
dence theory for the final output. The combined model LSTM model’s performance is compared with RNN and
predicts at most 50 hours in advance, and the results show MLP models, and the LSTM provides better results among
that the model accuracy reduces significantly over 25 hours. all the investigated methods based on the results. Hamshaw
Liu et al. (2019) use an LSTM model to forecast drinking et al. () use the Restricted Boltzmann Machine (RBM) to
water quality for up to 6 months. Zou et al. () develop classify the sediment-discharge curve in 14 categories from
an ensemble approach to predict water quality data, such as the 2D image of the suspended-sediment discharge plots
pH, DO, CODMn, and NH3-N. The approach based on from 600þ storm events. Finally, an LSTM based system is
using three LSTM models that different size interval data proposed to identify the characteristics of the water pollu-
feed each of them and the final prediction is a combination tants and trace its sources in the work of Wang et al.
of the results of three LSTM models. Banerjee et al. () (2019). In the system, a water quality cross-validation map
choose the indicators, namely dissolved oxygen and zooplank- is generated to identify pollutants and, based on defined
ton abundance, to reflect the water quality level of a reservoir. rules, track the pollutants to common industries.
An ANN model is proposed to model the selected indi-
cators in order to represent the water quality level. Yu et al.
() combine the LSTM model with Wavelet Mean Fusion WATER RESOURCES MANAGEMENT
and Wavelet Domain Threshold Denoising to simulate the
change of chlorophyll-a concentration in Dianchi Lake, Urban water systems are essential to modern cities. Effi-
China and use 15 water quality parameters as inputs, such cient operation of water treatment plants, wastewater

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
24 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

treatment plants, and conveyance networks require accu- the distributed water system as input and returns the poss-
rate modelling of these interconnected systems. Zhang ible locations of leakage points.
et al. (b) investigate the multiple models to simulate Various studies explore the strength of deep learning
and predict the water level of Combined Sewer Overflow powered modeling in water and wastewater treatment
(CSO) structure. In the study, the collected data from IoT procedures. Shi & Xu () develop a Stacked Denoising
is used separately with four different neural networks, AutoEncoders (SDAE) to predict a two-stage biofilm sys-
namely MLP, ANN with Wavelet, LSTM, and GRU, to com- tem’s performance. Nam et al. () propose a DQN-
pare the networks with each other. According to the results, based system to operate membrane bioreactor (MBR)
LSTM and GRU have good performances with respect to more efficiently. It aims to maximize the system’s energy effi-
others, but GRU has a quicker learning curve, fewer par- ciency while meeting stringent discharge qualities.
ameters, and simpler architecture. Despite these GoogLeNet (Szegedy et al. ) architecture is used to
advantages, the accuracy of GRU is slightly lower than the identify and classify microbeads in urban wastewater into
LSTM. Zhang et al. (a) predict the next hour’s waste- five categories based on microscopic images in the work
water flow to avoid sewer overflow using LSTM with the of Yurtsever & Yurtsever (). Harrou et al. () provide
traditional hydraulic model on the sewer system. This one- a case study using a DBM-SVM model to identify abnormal
hour prediction is tested in several scenarios, which are: 1- signals from the water features such as pH, conductivity, etc.
time step prediction at the 1-hr sampling frequency, 2-time in wastewater treatment plants. The results show that it is
steps prediction at the 30-min sampling frequency, 4-time possible to detect the abnormal conditions in order to
steps prediction at 15-min sampling frequency, and 6-time alert the system early based on the outcome of the DBM-
steps prediction at 10-min sampling frequency. The perform- SVM model. Mamandipoor et al. () develop an LSTM
ance of the LSTM based model is compared with Support model to monitor a wastewater treatment plant for detecting
Vector Regression (SVR) and Feed Forward Neural Net- faults during the oxidation and nitrification processes.
work (FFNN), and the LSTM based model has the best
accuracy in all scenarios. Zhang et al. (c) also predict
the next hour’s wastewater inflow for the wastewater treat- WEATHER
ment plant. The paper aims to identify which parts of the
sewer system have more free space and take action based Rainfall forecasting is one of the significant tasks in the
on the outcome. LSTM, NARX, Elman Neural Network domain of meteorology. Several techniques have been
are compared, and the LSTM model provides better results proposed to forecast rainfall with the help of statistics,
than other methods based on the results. Karimi et al. () machine learning, and deep learning. Zhang et al. ()
propose an LSTM model to forecast flow in sanitary sewer introduce an ensemble approach to forecast rainfall. In the
systems. It is claimed that accepting the groundwater as an first step, eight major meteorological factors are selected
additional input for the LSTM model increases the overall via correlation analysis between control forecast meteorolo-
accuracy of the task. gical factors and real-time rainfall. Then, samples are
In addition to the prediction of the wastewater level or divided into four categories by K-means clustering. The
flow, some studies aim to detect conveyance network con- LSTM based model is fed by each cluster, and outputs are
ditions. Xu et al. () aim to detect abnormal working combined to reach the final prediction. Weber et al. ()
conditions in the water supply network. Besides the detec- develop a CNN-based surrogate model for one of the
tion of abnormal conditions, pressure in the water supply global climate model CanESM2 (Arora et al. ). The
network is predicted. An LSTM model is developed to CNN model is fed by 97 years of monthly precipitation
achieve these goals. The performance of the LSTM model output from CanESM2, and the model preserves its perform-
outperforms the traditional prediction models, such as ance even when the forecast length is expanded to 120
SVM. Zhou et al. () use an ANN model to identify months. According to the paper, the accuracy of the
burst locations in a water distribution network. The model model can be increased by deeper networks. Poornima &
takes the pressure data as input and returns one or several Pushpalatha () investigate an LSTM model to forecast
possible pipes, which can be the location of the burst. the rainfall with the help of 34 years of rainfall data. In
Fang et al. () focus on detecting multiple leakage the paper, the LSTM model results are compared with mul-
points in a water distribution network with a CNN based tiple methods, such as Holt-Winters and ARIMA. Tang et al.
model. The model accepts the pressure measurements of () introduce a DNN model to predict rain and snow

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
25 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

rates at high altitudes. In the paper, passive microwave, methods, including linear regression, nearest neighbor,
infrared and environmental data are trained to the reference random forest, and DNN. Wang et al. (c) develop an
precipitation data sets, which are obtained by two space- RNN model to perform statistical downscaling on tempera-
borne radars for the estimations. The results of the DNN ture and precipitation in order to improve the accuracy of
model are compared with many methods, such as the God- hydrological models. The RNN model provides better accu-
dard Profiling Algorithm (GPROF). The experiment results racy than the compared methods, such as ANN, for the
show that the DNN model is capable of predicting snow extreme temperature and precipitation downscaling based
and rain rate more accurately than other tested methods at on the evaluation of downscaled data on the Soil and
high altitudes. Water Assessment Tool (SWAT) model.
Some studies specifically focus on the improvement of In meteorology and remote sensing, cloud or cloud
quantitive precipitation estimation accuracy, alongside shadow detection has received attention in the recent
precipitation nowcasting. Since the prediction of precipi- literature. Because U-net (Ronneberger et al. ), a CNN
tation generally is a time series problem, the usage of model for biomedical image segmentation, and Segnet
LSTM architecture is common on this task. Wu et al. (Badrinarayanan 2017), a convolutional encoder-decoder
() design a fusion model, which is a combination of architecture for image segmentation, provide successful
CNN and LSTM, to improve quantitative precipitation esti- results in their domain, many researchers use those
mation accuracy. The proposed model uses satellite data, models as a base model for their works. Jeppesen et al.
rain gauge data, and thermal infrared images. The CNN () use the U-net to detect clouds in satellite imagery.
part of the model extracts the spatial characteristics of the The model is trained and tested with Landsat 8 Biome and
satellite, rain gauge, and thermal infrared data, where The SPARCS datasets. Wieland et al. () aim to segment
LSTM part of the model handles the time dependencies of cloud, shadow, snow/ice, water, and land in multi-spectral
the provided data. The performance of the CNN-LSTM satellite images. The U-net based model is proposed and
model is better than the comparative models, such as trained with Landsat datasets to segment images into five
CNN, LSTM, and MLP. Yan et al. () use a CNN categories. According to results, contrast and brightness aug-
model to forecast short-term precipitation with the help of mentations of the training data improve the segmentation
radar reflectance images for a local area in China. As a data- accuracy, alongside adding shortwave-infrared bands.
set, the radar reflection images and the corresponding Zhang et al. (b) applied the U-net based model on the
precipitation values for one hour are collected. The model red, green, blue, and infrared waveband images from the
takes the images as inputs and returns the forecast value Landsat-8 dataset for cloud detection. LeGall-5/3 wavelet
for one-hour precipitation. The CNN model contains transform is used on the dataset to accelerate the model
residual links between the layers, which increase the effi- and make it feasible to implement on-board on satellite.
ciency of the model. Chen et al. () focus on Chai et al. () propose a CNN model based on Segnet
precipitation nowcasting using a ConvLSTM model. The to detect clouds and cloud shadow in Landsat imagery.
model accepts the radar echo data in order to forecast 30 Thunderstorms and typhoons are one of the extreme
or 60 minutes of precipitation value. According to natural disasters that can cause massive damages. Some
results, using the customized multisigmoid loss function studies focus on the prediction of those hazardous natural
and group normalization provides better performance than events to take early actions to minimize the damage.
ConvLSTM with classical loss functions, such as cross- Ukkonen & Mäkelä () aim to predict the occurrence
entropy loss function, and conventional extrapolation of thunderstorms from parameters related to instability, inhi-
methods. bition, and moisture mainly. A DNN model is trained with
Statistical downscaling methods often provide more lightning data and a high-resolution global reanalysis. Var-
accurate precipitation estimation than using raw precipi- ious regions, such as Sri Lanka and Europe, are used as
tation values in the models. test areas for the model. Many valuable findings are pro-
Deep learning methods can be used as statistical vided related to the correlation between thunderstorm
downscaling methods to improve the accuracy of the occurrence and parameters specific to regions. Kim et al.
tasks. Pan et al. () propose a CNN model as a statistical () aim to find similarities between the typhoon and a
downscaling method (SD) for daily precipitation prediction. typhoon from the past for helping to mitigate the effect of
The method is tested with 14 geogrid points in the U.S., and the typhoon. In the study, a DNN model is used to encode
SD results from the CNN model outperform other tested the typhoon event by typhoon parameters, such as route,

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
26 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

pressure, and moving speed. The model returns the typhoon training process, Structural Similarity (SSIM) and multi-
events’ similarity to the historical ones, which provide scale SSIM are used to obtain better results.
insights for officials to take early action. During the work,
a database is created for 189 typhoons that occurred
between 1950 and 2017. Wei & Cheng () aim to predict UNCLASSIFIED STUDIES
wind speed and wave height of a typhoon with the help of an
RNN model, namely TSWP. The TSWP model predicts Some literature included in this review did not fit within any
wind speed first, followed by wind height for 1–6 hours in of the defined categories in this review. Summaries of these
the future. The results of the TSWP model outperform the papers are provided in this section. Many of the papers in
comparative methods, such as MLP, DNN, and Logistic this section apply deep learning methods to ocean pro-
Regression. cesses. In coastal hydraulics Kang () used an improved
In addition to the aforementioned studies, various CNN on images to classify and monitor waves. Also in
works have been published for different meteorological coastal hydraulics, Kim et al. () used a number of
tasks. Su et al. () use the pyramid delaminating tech- deep neural networks with coastal video imagery to develop
nique to generate the global optical flow field. A a framework to track nearshore waves. Kylili et al. () use
ConvLSTM model takes the RGB image and generated a CNN to identify floating plastic debris in ocean images.
flow field in order to improve the forecast accuracy of Further, Wang et al. (2019) used a CNN architecture with
echo position and intensity. The generation, dissipation, a satellite radar product to classify ocean surface roughness
and merging of convective cells were also better identified into ten geophysical phenomena. Li et al. () also used a
in comparison to other classical methods. Weyn et al. CNN to classify hurricane damage from post-event aerial
() investigate to forecast weather at 500-hPa geopoten- imagery.
tial height by a CNN model. The paper aims to use The remaining three papers in this section apply deep
historical gridded reanalysis data in the model without expli- learning in a variety of disciplines. Haklidir & Haklidir
cit knowledge about the physical process. The CNN model () use a DNN to predict the temperature of geothermal
produces promising results for capturing the climatology springs and wells given hydrogeochemical data, including
and annual variability of 500 - hPa heights and predicts rea- chemical concentrations. Amirkolaee & Arefi ()
listic atmospheric states for 14 days. However, the CNN implemented a very deep CNN structure to estimate a digital
model still could not perform as good as an operational elevation model from single airborne or spaceborne images.
weather model. He et al. (a) propose an LSTM model Finally, Rohmat et al. () developed and embedded a
to classify periods as rainy or dry by microwave links. DNN into a GIS-based basin-scale decision support system
Scher & Messori () aim to predict the uncertainty of a to assess the impacts of best management practices (BMP.)
forecast based on past forecasts and their occurrence rate Based on geographic information and potential BMP, the
with a CNN model. The CNN model takes the atmospheric DNN returns multiple outputs which describe net flow sep-
fields at three different heights as inputs and returns the aration between groundwater recharge, groundwater return
scalar value representing the predictability (or the uncer- flow, and overland flow. This project demonstrates the
tainty) of precipitation forecast. Klampanos et al. () potential of deep learning to integrate into decision
study the relationship between nuclear events and weather. making on large-scale water resources projects.
A model with Autoencoder and CNN is used for rapid
source estimation during radiological releases. The model
clusters weather events first and looks over their corre- KEY ISSUES AND CHALLENGES
lations with nuclear events. Chen et al. () design a
GAN model based on SRGAN (Ledig et al. ), which is Deep learning captures the non-linear complex represen-
a GAN architecture for single image super-resolution, to tations with datasets, making it a viable and efficient
improve the resolution of radar echo images of weather option in predictive model creation. In machine learning lit-
radar systems in order to increase the accuracy of the erature capturing the representation is known as
tasks that accepts echo images as inputs. Tran & Song representation learning (Goodfellow et al. ). Represen-
() design a ConvRNN model for the radar echo extrapol- tation learning relies on artificial neural networks’ ability
ation task. The model uses multiple (five) satellite images in acting as a universal approximator (Cybenko ;
and predicts 10 steps ahead at the pixel level. During the Hornik et al. ; Leshno et al. ) meaning ANNs

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
27 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

with only one hidden layer, theoretically, could represent governmental and agency-collected water data and its use
any function. The drawback regarding this ability is that in deep learning.
the task in hand might need a hidden layer that is too We consider the fundamental understanding of deep
large to still be feasible to be trained and executed. Adding learning within the literature as another problematic issue
new hidden layers to neural networks comes into play in emerging from the reviews made in this study. A common
order to cover this negation. Consequently, one important mistake regards what deep learning is and what deep learn-
principle of representation learning is that, with minimal ing is not. Most papers seem to interpret deep learning as a
data engineering, the datasets should be fed to the neural specific technique. However, deep learning is simply a broad
network and let the neural network decide which features term for various machine learning algorithms centered upon
within the dataset are important towards the goal of repre- ANNs. Various studies reviewed here claim that they
senting that dataset. Even though this should be the case, employ deep learning, yet they only take advantage of
the literature in the water domain does not widely apply traditional ANN approaches. Most of the time this utiliz-
this principle. We attribute this to the fact that datasets are ation doesn’t surpass the extent of a study that employs
neither extensive enough in terms of the number of given conventional statistical modelling. This phenomenon raises
data, nor have vast spatial and/or temporal coverage. questions whether these studies attempt to exploit the
As opposed to fields like computer vision and natural keywords, using the term ‘deep learning’ to take advantage
language processing, the water field lacks high quality, col- of the current scientific zeitgeist. We observe that these
lected, curated, labeled, and published datasets that are motives rarely result in work that forwards the deep learning
used for benchmarking and method development. We ident- literature. We particularly note the poor practice of
ify this lack of benchmarking datasets as a key challenge ascribing work ‘deep learning’ while not employing rep-
(Ebert-Uphoff et al. ) which slows the state-of-the- resentation learning principles.
science deep learning applications in water domain. Most This comprehensive review identified that literature at
studies reviewed herein acquire datasets from governmental the intersection of deep learning and water do not cover
agencies depending on their needs. Though potentially con- the used methods in detail. Most of the studies appear
venient for a contained effort of a single work, each one-off reluctant to give model/architecture details that are vital
dataset dampens the speed of improvement of the state of to reproduce the proposed training pipeline. Combined
the science in this field. If researchers had the opportunity with the aforementioned dataset problems cause consistent
to build on a widely-accepted dataset in their fields, they barriers to reproducibility of work. This overshadows the
would be able to improve the accuracy of their models by reported accuracy of studies and slows advancement in
taking advantage of previous models created using the same the deep learning powered water field.
dataset. This collaboration around common datasets and In contrast, some studies discuss unnecessary details
models would open opportunities in the field, such as paving over and over again in their manuscript. Discussing how
the way for their real-time usage. However, as we don’t have the most efficient number of hidden layers or hyper-
many benchmarking datasets, many research groups all parameters are found by trial and error for many paragraphs
around the globe run similar networks on custom datasets without a hint of intuition, or discussing the optimal batch
they acquired with limited scientific interaction. size for many paragraphs do not appear as the authors use
One other problem with the data provided by authorities pertinent deep learning theory. Not sharing intuition regard-
is that they are dispersed among different agencies and they ing technical choices also prevents the papers to reach their
occasionally have mismatches in temporal and/or spatial goal. There are many papers utilizing LSTM networks, but a
coverage. Even though the data provided extends to many few discuss the rationale behind their decision. One would
years before covering decades, one might need to access var- at least expect a paper to describe why they choose LSTM
ious databases created by several different government networks over GRU networks in their setting. While this
agencies in order to build a dataset. Further, the period of instance needs a comparison of two similar RNN implemen-
record commonly differs across agencies and areas which tations, it should also be noted that networks like GANs and
causes the data acquisition process time-consuming and DQNs find a very limited place for themselves in the litera-
sometimes inconclusive. Further complicating data acqui- ture. Furthermore, researchers use networks like ELM,
sition, water data are suppressed due to their military ENN, and NAR in various tasks without sharing their rela-
significance in some strategic instances and contributes to tive success to more complex architectures widely-used in
limit progress in the field. We support growing access to state of the art deep learning models.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
28 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

ETHICS IN DL APPLICATIONS task the algorithm is in service of, feedback on algorithm per-
formance, and rigorous assessment procedures.
Broadly, ethics are concerned with describing the appropri- The discussions of ethics in the context of DL are part of
ateness of an act; whether something is ‘right’ or ‘wrong’ to a continued conversation of ethics generally. Questions of
do. Meanwhile, DL is a tool which can be used to produce how to treat those in our communities, how to treat the
algorithms which automate prediction and decision- environment, what is the ‘right’ priority, or priorities, are
making. Thus, most ethical concerns of DL derive from not questions exclusive to the AI/DL domain. Rather, the
the central question: What will the DL application do? new scale and speed provided by DL require the inspection
The ethical considerations of DL in water are no exception. of age-old ethical questions in a new light.
Take, for example, a DL application that predicts stream-
flow within an urban catchment. By appearances, the act
of streamflow prediction alone lacks any ethical character.
However, if the same tool is then used to make decisions RECOMMENDATIONS AND CONCLUSIONS
in disaster mitigation (Demir et al. ) or public planning,
many questions with ethical dimensions arise (Ewing & This paper provides a comprehensive review of the recent
Demir ). What data were, and were not, used to develop application of deep neural networks as novel solutions to
the model? What biases exist in the dataset? How do these tackle hydrological tasks and challenges. A total of 129 pub-
biases affect decision-making and human lives, and do the lications were systematically selected for rigorous review
decisions reveal any discriminating behaviors? From this and analysis as grouped by their application area. Based
example, it is clear that the primary ethical considerations on the statistical meta-analysis of journal publications
for the application of DL in water should be concerned dating between January 2018 and March 2020, it was
with how entities – people, the environment, communities empirically observed that the average number of deep learn-
– will be affected by DL in decision-making workflows. ing applications per month in the water sector steadily
Though powerful, these DL tools simultaneously expand increased in an exponential fashion. Further, the rapidly
the reach and speed of decision-making, while also stripping increasing body of work from these publications show
away layers of context that would possibly be relevant to a deep learning’s potential in a diversity of water sector appli-
humans’ decision-making process. However, many of the ethi- cations. Key issues and challenges that may constitute
cal decisions themselves in the water field remain unchanged setbacks and hindrances of deep learning adoption in
and are primarily distributional; who receives water services, the water industry have been identified and reported
and their level of quality, and what level, and for whom, of accompanied by recommendations to persevere in spite of
risk is acceptable. These persistent ethical water dilemmas the logistical, computational, expertise-related, and data-
must be resolved within the new paradigm of DL in water. related challenges to its principled adoption. Based on the
As revealed by this literature review, few DL applications extent of this review and the broad spectrum of application
in the water literature include decision-making components. areas, we anticipate the water sector will continue to incor-
One paper, Rohmat et al. (), reviewed in this paper expli- porate deep learning at an accelerating rate and deep
citly states that their DL tool is integrated into a decision learning will play a key role in the future of water.
support system. This lack of attention to DL in service of Deep learning-powered technologies opened up a
decision-making in water academia presents an opportunity plethora of application and research opportunities to revolu-
for a new line of research. It is also an opportunity to incorpor- tionize hydrological science and workflow. In a bird’s eye
ate the work of other fields early, such as the well-documented view fashion, key areas of innovation for future research
ethical concerns stemming from DL decision- and recommen- include:
dation-engines in social and civil applications (such as 1. Automated Forecasting: As this review outlined, the
policing, criminal justice, and self-driving cars (Angwin et al. majority of current deep learning applications in the
; O’Neil ).) Further, a proliferation of AI Ethics frame- hydrological domain focus on forecasting of numerous
works have been developed, reviewed here (Hagendorff ; variables (e.g. water level, discharge) given the problem’s
Jobin et al. ), as well as numerous guidelines for applying suitability for machine learning. In the future, efforts
and assessing algorithms in social and civil domains (O’Reilly can be coordinated between agencies and research
; Rahwan ; Reisman Schultz Crawford & Whittaker organizations to collaboratively develop complementary
). These guidelines stress a deep understanding of the models that will yield actionable and reliable

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
29 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

information. These models can be maintained and pow- which can allow the stakeholders of the water domain
ered by a stream of real-time data influx to constitute to innovate novel applications with existing or low-cost
the future of decision-making systems and geographical sensor networks. As a tangible example, a camera-
information systems. equipped river monitoring sensor can employ deep learn-
2. Published Datasets: Lack of deep learning ready datasets ing to analyze pictures on the edge to detect any foreign
within the water field was stressed in the previous sec- objects (e.g. tree, human) on the river, and transmit only
tion. The main problem caused by this absence of many the useful information to a centralized system.
datasets is that the research community does not build 5. Intelligent Assistants: The massive amount of environ-
upon previous work in terms of constructing better mental and hydrological data makes it challenging to
neural network architectures and moving the state of efficiently and effectively extract required knowledge in
art to the next iteration. This inference is supported by a timely manner. Manual efforts are often needed to ana-
the fact that among the 30 papers related to flooding lyze comprehensive raw data for decision-making
reviewed in this study there are only a few that use a pre- purposes. As a solution, intelligent assistants serve as
viously curated, labeled dataset. The result is many voice-enabled knowledge engines that can interpret a
papers are published that achieve the same task with natural language question, apply human-like reasoning,
almost identical methods but different data. This absence and extract the desired factual response out of curated
implicitly causes redundancy in the field. We believe if data. These assistants can be integrated into various com-
more studies focus on creating benchmark datasets that munication channels for easy access including web
are open to researchers, both cumulativeness of the systems, messaging applications (e.g. Skype, Messenger,
science would be satisfied and deep learning powered SMS), personal assistants (e.g. Google Assistant, Apple
modeling in water resources research would go further Siri), home automation devices (e.g. Google Home,
in terms of generic applicability. Amazon Alexa), augmented and virtual reality systems
3. AI as a Service: As the popularity and usefulness of arti- (e.g. HoloLens, Magic Leap, Oculus Quest), and
ficial intelligence tools increase, a new research area automated workflow systems. The voice-enabled com-
came to prominence in the computer science field. This munication and immediate access to knowledge can
area is focused on developing generalized and centra- facilitate hydrological research as well as natural disaster
lized web frameworks that can readily provide the preparedness and response (Sermet & Demir a).
means to develop custom AI solutions through Plat- 6. Virtual and Augmented Reality: Incorporation of deep
form-as-a-Service (PaaS) systems. These systems hold learning with virtual and augmented reality environ-
great potential for the hydrological community as they ments provides a prominent research area due to its
allow developers to focus on designing and managing immersive nature that allows effective analysis of com-
intelligent applications without the burden of infrastruc- plex environmental phenomena that is not feasible to
ture maintenance and adjustment of computing be orchestrated in real-life. For example, deep learning
resources. They can provide intuitive graphical user inter- can power realistic flood simulations while mimicking
faces to allow the development of hydrological deep human behavior to train first-responders, aid decision-
learning applications by connecting predefined and makers, and educate the public (Sermet & Demir
custom models with provided datasets. b). Another use case may be guiding on-site person-
4. Edge Computing: The main propeller in the creation of nel that may require expertise, such as sensor
smart applications is the consistent and diverse data flux. maintenance, structural renovation, and field exper-
However, as the frequency and types of data resources iments, through heads-up displays and deep learning-
expand, a centralized approach to collect and analyze powered recognition and decision support applications
data simply may not be viable for multivariate tasks. Fur- (Sermet & Demir ).
thermore, the costs associated with the transfer of large
data from distributed sensors are not trivial and may dis-
courage stakeholders to increase sensor coverage and
data reporting interval. As a solution, edge computing DATA AVAILABILITY STATEMENT
offers a new perspective to process the data on the
sensor. There is extensive research on utilizing deep learn- All relevant data are included in the paper or its Supplemen-
ing for the internet of things through edge computing, tary Information.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
30 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Journal of the Indian Society of Remote Sensing 47 (11),

REFERENCES
1949–1958.
Bui, D. T., Hoang, N. D., Martínez-Álvarez, F., Ngo, P. T. T., Hoa,
Abdi, G., Samadzadegan, F. & Reinartz, P.  Deep learning
P. V., Pham, T. D., Samui, P. & Costache, R. a A novel
decision fusion for the classification of urban remote sensing
deep learning neural network approach for predicting
data. Journal of Applied Remote Sensing 12 (1), 016038.
flash flood susceptibility: a case study at a high frequency
Agliamzanov, R., Sit, M. & Demir, I.  Hydrology@ home: a
tropical storm area. Science of The Total Environment 701,
distributed volunteer computing framework for hydrological
134413.
research and applications. Journal of Hydroinformatics
Bui, Q. T., Nguyen, Q. H., Nguyen, X. L., Pham, V. D., Nguyen,
22 (2), 235–248.
H. D. & Pham, V. M. b Verification of novel integrations
Amirkolaee, H. A. & Arefi, H.  Convolutional neural network
of swarm intelligence algorithms into deep learning neural
architecture for digital surface model estimation from
network for flood susceptibility mapping. Journal of
single remote sensing image. Journal of Applied Remote
Hydrology 581, 124379.
Sensing 13 (1), 016522.
Cao, F., Yang, Z., Ren, J. & Ling, B. W. K.  Convolutional
Angwin, J., Larson, J., Mattu, S. & Kirchner, L.  Machine Bias.
neural network extreme learning machine for effective
ProPublica, May, 23, p. 2016.
classification of hyperspectral images. Journal of Applied
Antonetti, M. & Zappa, M.  How can expert knowledge
Remote Sensing 12 (3), 035003.
increase the realism of conceptual hydrological models?
Carson, A., Windsor, M., Hill, H., Haigh, T., Wall, N., Smith, J.,
A case study based on the concept of dominant runoff
Olsen, R., Bathke, D., Demir, I. & Muste, M.  Serious
process in the Swiss Pre-Alps. Hydrology & Earth System
gaming for participatory planning of multi-hazard mitigation.
Sciences 22 (8).
International Journal of River Basin Management 16 (3),
Ardabili, S., Mosavi, A., Dehghani, M. & Várkonyi-Kóczy, A. R.
379–391.
 Deep learning and machine learning in hydrological
Chai, D., Newsam, S., Zhang, H. K., Qiu, Y. & Huang, J. 
processes climate change and earth systems a systematic
Cloud and cloud shadow detection in Landsat imagery based
review. In International Conference on Global Research and
on deep convolutional neural networks. Remote Sensing of
Education. Springer, Cham, pp. 52–62.
Environment 225, 307–316.
Arora, V. K., Scinocca, J. F., Boer, G. J., Christian, J. R., Denman,
Chen, H., Zhang, X., Liu, Y. & Zeng, Q.  Generative
K. L., Flato, G. M., Kharin, V. V., Lee, W. G. & Merryfield,
adversarial networks capabilities for super-resolution
W. J.  Carbon emission limits required to satisfy future
reconstruction of weather radar echo images. Atmosphere
representative concentration pathways of greenhouse gases.
10 (9), 555.
Geophysical Research Letters 38 (5).
Chen, L., Cao, Y., Ma, L. & Zhang, J.  A deep learning-based
Badrinarayanan, V., Kendall, A. & Cipolla, R.  Segnet: a deep
methodology for precipitation nowcasting with radar. Earth
convolutional encoder-decoder architecture for image
and Space Science 7 (2), e2019EA000812.
segmentation. IEEE Transactions on Pattern Analysis and
Cho, K., Van Merriënboer, B., Bahdanau, D. & Bengio, Y.  On
Machine Intelligence 39 (12), 2481–2495.
the Properties of Neural Machine Translation: Encoder-
Bai, Y., Bezak, N., Sapač, K., Klun, M. & Zhang, J.  Short-term
Decoder Approaches. arXiv preprint arXiv:1409.1259.
streamflow forecasting using the feature-enhanced regression
Cybenko, G.  Approximation by superpositions of a sigmoidal
model. Water Resources Management 33 (14), 4783–4797.
function. Mathematics of Control, Signals and Systems 2 (4),
Banerjee, A., Chakrabarty, M., Rakshit, N., Bhowmick, A. R. &
303–314.
Ray, S.  Environmental factors as indicators of dissolved
Damavandi, H. G., Shah, R., Stampoulis, D., Wei, Y., Boscovic, D.
oxygen concentration and zooplankton abundance: deep
& Sabo, J.  Accurate prediction of streamflow using long
learning versus traditional regression approach. Ecological
short-term memory network: a case study in the Brazos River
Indicators 100, 99–117.
Basin in Texas. International Journal of Environmental
Barzegar, R., Aalami, M. T. & Adamowski, J.  Short-term
Science and Development 10 (10), 294–300.
water quality variable prediction using a hybrid CNN–LSTM
Demir, I. & Szczepanek, R.  Optimization of river network
deep learning model. Stochastic Environmental Research and
representation data models for web-based systems. Earth and
Risk Assessment 1–19.
Space Science 4 (6), 336–347.
Bengio, Y., Frasconi, P. & Simard, P.  The problem of learning
Demir, I., Yildirim, E., Sermet, Y. & Sit, M. A.  FLOODSS:
long-term dependencies in recurrent networks. In IEEE
Iowa flood information system as a generalized flood
International Conference on Neural Networks. IEEE,
cyberinfrastructure. International Journal of River Basin
pp. 1183–1188.
Management 16 (3), 393–400.
Bhola, P. K., Nair, B. B., Leandro, J., Rao, S. N. & Disse, M. 
Demiray, B. Z., Sit, M. & Demir, I.  D-SRGAN: DEM Super-
Flood inundation forecasts using validation data generated
Resolution with Generative Adversarial Network. arXiv
with the assistance of computer vision. Journal of
preprint arXiv:2004.04788.
Hydroinformatics 21 (2), 240–256.
Ebert-Uphoff, I., Thompson, D. R., Demir, I., Gel, Y. R., Karpatne,
Bhosle, K. & Musande, V.  Evaluation of deep learning CNN
A., Guereque, M., Kumar, V., Cabral-Cano, E. & Smyth, P.
model for land use land cover classification and crop
 A vision for the development of benchmarks to bridge
identification using hyperspectral remote sensing images.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
31 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

geoscience and data science. In 17th International Workshop He, B., Liu, X., Hu, S., Song, K. & Gao, T. a Use of the C-band
on Climate Informatics. microwave link to distinguish between rainy and dry periods.
Elman, J. L.  Finding structure in time. Cognitive Science Advances in Meteorology, 2019.
14 (2), 179–211. He, X., Luo, J., Zuo, G. & Xie, J. b Daily runoff forecasting
Esteva, A., Robicquet, A., Ramsundar, B., Kuleshov, V., DePristo, using a hybrid model based on variational mode
M., Chou, K., Cui, C., Corrado, G., Thrun, S. & Dean, J.  decomposition and deep neural networks. Water Resources
A guide to deep learning in healthcare. Nature Medicine Management 33 (4), 1571–1590.
25 (1), 24–29. Heaton, J. B., Polson, N. G. & Witte, J. H.  Deep learning for
Evora, N. D. & Coulibaly, P.  Recent advances in data-driven finance: deep portfolios. Applied Stochastic Models in
modeling of remote sensing applications in hydrology. Business and Industry 33 (1), 3–12.
Journal of Hydroinformatics 11 (3–4), 194–201. Hinton, G. E.  Training products of experts by minimizing
Ewing, G. & Demir, I.  An ethical decision-making contrastive divergence. Neural Computation 14 (8),
framework with serious gaming: smart water case study on 1771–1800.
flooding. EarthArXiv. doi:10.31223/osf.io/y873 g. Hinton, G. E.  Deep belief networks. Scholarpedia 4 (5), 5947.
Fang, Q., Zhang, J., Xie, C. & Yang, Y.  Detection of multiple Hochreiter, S. & Schmidhuber, J.  Long short-term memory.
leakage points in water distribution networks based on Neural Computation 9 (8), 1735–1780.
convolutional neural networks. Water Supply 19 (8), Hornik, K., Stinchcombe, M. & White, H.  Multilayer
2231–2239. feedforward networks are universal approximators. Neural
Fu, T., Ma, L., Li, M. & Johnson, B. A.  Using convolutional Networks 2 (5), 359–366.
neural network to identify irregular segmentation objects Hrnjica, B. & Bonacci, O.  Lake level prediction using feed
from very high-resolution remote sensing imagery. Journal of forward and recurrent neural networks. Water Resources
Applied Remote Sensing 12 (2), 025010. Management 33 (7), 2471–2484.
Gautam, A., Sit, M. & Demir, I.  Realistic River Image Hu, R., Fang, F., Pain, C. C. & Navon, I. M.  Rapid spatio-
Synthesis Using Deep Generative Adversarial Networks. temporal flood prediction and uncertainty quantification
arXiv preprint arXiv:2003.00826. using a deep learning method. Journal of Hydrology 575,
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, 911–920.
D., Ozair, S., Courville, A. & Bengio, Y.  Generative Huang, G. B., Zhu, Q. Y. & Siew, C. K.  Extreme learning
adversarial nets. In Advances in Neural Information machine: theory and applications. Neurocomputing 70 (1–3),
Processing Systems. pp. 2672–2680. 489–501.
Goodfellow, I., Bengio, Y. & Courville, A.  Deep Learning. Islam, Z.  A Review on Physically Based Hydrologic Modeling.
MIT press. University of Alberta, Edmonton, AB, Canada. doi:10.13140/
Grossman, D., Buckley, N. & Doyle, M.  Data Intelligence for 2.1.4544.5924.
21st Century Water Management: A Report From the 2015 Isola, P., Zhu, J. Y., Zhou, T. & Efros, A. A.  Image-to-image
Aspen-Nicholas Water Forum. Aspen-Nicholas Water translation with conditional adversarial networks. In
Forum. Available from https://www.aspeninstitute.org/ Proceedings of the IEEE Conference on Computer Vision and
publications/data-intelligence-21st-century-water- Pattern Recognition. pp. 1125–1134.
management-report-2015-aspen-nicholas-water-forum/ Jadidoleslam, N., Mantilla, R., Krajewski, W. F. & Cosh, M. H.
(accessed 31 May 2020).  Data-driven stochastic model for basin and sub-grid
Hagendorff, T.  The Ethics of AI Ethics–An Evaluation of variability of SMAP satellite soil moisture. Journal of
Guidelines. arXiv preprint arXiv:1903.03425. Hydrology 576, 85–97.
Haklidir, F. S. T. & Haklidir, M.  Prediction of reservoir Jaiswal, R. K., Ali, S. & Bharti, B.  Comparative evaluation of
temperatures using hydrogeochemical data, Western conceptual and physical rainfall–runoff models. Applied
Anatolia geothermal systems (Turkey): a machine learning Water Science 10 (1), 1–14.
approach. Natural Resources Research 1–14. Jeong, J. & Park, E.  Comparative applications of data-driven
Hamshaw, S. D., Dewoolkar, M. M., Schroth, A. W., Wemple, models representing water table fluctuations. Journal of
B. C. & Rizzo, D. M.  A new machine-learning approach Hydrology 572, 261–273.
for classifying hysteresis in suspended-sediment discharge Jeong, J., Park, E., Chen, H., Kim, K. Y., Han, W. S. & Suk, H. 
relationships using high-frequency monitoring data. Water Estimation of groundwater level based on the robust training
Resources Research 54 (6), 4040–4058. of recurrent neural networks using corrupted data. Journal of
Harrou, F., Dairi, A., Sun, Y. & Senouci, M.  Statistical Hydrology 582, 124512.
monitoring of a wastewater treatment plant: a case study. Jeppesen, J. H., Jacobsen, R. H., Inceoglu, F. & Toftegaard, T. S.
Journal of Environmental Management 223, 807–814.  A cloud detection algorithm for satellite imagery based
Haskins, G., Kruger, U. & Yan, P.  Deep learning in medical on deep learning. Remote Sensing of Environment 229,
image registration: a survey. Machine Vision and 247–259.
Applications 31 (1), 8. Jiang, W.  Object-based deep convolutional autoencoders for
He, K., Zhang, X., Ren, S. & Sun, J.  Deep Residual Learning high-resolution remote sensing image classification. Journal
for Image Recognition. arXiv preprint arXiv:1512.03385. of Applied Remote Sensing 12 (3), 035002.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
32 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Jiang, Z., Mallants, D., Peeters, L., Gao, L., Soerensen, C. & Kratzert, F., Klotz, D., Shalev, G., Klambauer, G., Hochreiter, S. &
Mariethoz, G.  High-resolution paleovalley classification Nearing, G. b Towards learning universal, regional, and
from airborne electromagnetic imaging and deep neural local hydrological behaviors via machine learning applied
network training using digital elevation model data. to large-sample datasets. Hydrology & Earth System Sciences
Hydrology and Earth System Sciences 23 (6), 2561–2580. 23 (12).
Jin, B., Ye, P., Zhang, X., Song, W. & Li, S.  Object-oriented Krizhevsky, A., Sutskever, I. & Hinton, G. E.  Imagenet
method combined with deep convolutional neural networks classification with deep convolutional neural networks.
for land-use-type classification of remote sensing images. In Advances in Neural Information Processing Systems.
Journal of the Indian Society of Remote Sensing 47 (6), pp. 1097–1105.
951–965. Kroupi, E., Kesa, M., Navarro-Sánchez, V. D., Saeed, S., Pelloquin,
Jobin, A., Ienca, M. & Vayena, E.  Artificial Intelligence: The C., Alhaddad, B., Moreno, L., Soria-Frisch, A. & Ruffini, G.
Global Landscape of Ethics Guidelines. arXiv preprint  Deep convolutional neural networks for land-cover
arXiv:1906.11668. classification with Sentinel-2 images. Journal of Applied
Kabir, S., Patidar, S. & Pender, G.  Investigating capabilities of Remote Sensing 13 (2), 024525.
machine learning techniques in forecasting stream flow. In: Kruger, A., Krajewski, W. F., Niemeier, J. J., Ceynar, D. L. &
Proceedings of the Institution of Civil Engineers-Water Goska, R.  Bridge-mounted river stage sensors (BMRSS).
Management. Thomas Telford Ltd, Vol. 173, No. 2, pp. 69–86. IEEE Access 4, 8948–8966.
Kang, L.  Wave monitoring based on improved convolution Kumar, D., Singh, A., Samui, P. & Jha, R. K.  Forecasting
neural network. Journal of Coastal Research 94 (sp1), monthly precipitation using sequential modelling.
186–190. Hydrological Sciences Journal 64 (6), 690–700.
Kao, I. F., Zhou, Y., Chang, L. C. & Chang, F. J.  Exploring a Kylili, K., Kyriakides, I., Artusi, A. & Hadjistassou, C. 
long short-term memory based encoder-decoder framework Identifying floating plastic marine debris using a deep
for multi-step-ahead flood forecasting. Journal of Hydrology learning approach. Environmental Science and Pollution
124631. Research 26 (17), 17091–17099.
Karimi, H. S., Natarajan, B., Ramsey, C. L., Henson, J., Tedder, Laloy, E., Hérault, R., Jacques, D. & Linde, N.  Training-image
J. L. & Kemper, E.  Comparison of learning-based based geostatistical inversion using a spatial generative
wastewater flow prediction methodologies for smart sewer adversarial neural network. Water Resources Research 54 (1),
management. Journal of Hydrology 577, 123977. 381–406.
Kim, Y. J., Yura, E., Kim, T. W. & Yoon, J. S.  Development of LeCun, Y.  Generalization and network design strategies.
disaster prevention system based on deep neural network Connectionism in Perspective 19, 143–155.
using deep learning with dropout. Journal of Coastal LeCun, Y., Bengio, Y. & Hinton, G.  Deep learning. Nature
Research 91 (sp1), 186–190. 521 (7553), 436–444.
Kim, J., Kim, J., Kim, T., Huh, D. & Caires, S.  Wave-tracking Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta,
in the surf zone using coastal video imagery with deep neural A., Aitken, A., Tejani, A., Totz, J., Wang, Z. & Shi, W.  Photo-
networks. Atmosphere 11 (3), 304. realistic single image super-resolution using a generative
Klampanos, I. A., Davvetas, A., Andronopoulos, S., Pappas, C., adversarial network. In: Proceedings of the IEEE Conference on
Ikonomopoulos, A. & Karkaletsis, V.  Autoencoder- Computer Vision and Pattern Recognition. pp. 4681–4690.
driven weather clustering for source estimation during Lee, S. & Lee, D.  Improved prediction of harmful algal
nuclear events. Environmental Modelling & Software 102, blooms in four Major South Korea’s rivers using deep
84–93. learning models. International Journal of Environmental
Kopp, M., Tuo, Y. & Disse, M.  Fully automated snow depth Research and Public Health 15 (7), 1322.
measurements from time-lapse images applying a Leshno, M., Lin, V. Y., Pinkus, A. & Schocken, S.  Multilayer
convolutional neural network. Science of The Total feedforward networks with a nonpolynomial activation
Environment 697, 134213. function can approximate any function. Neural Networks
Krajewski, W. F., Ceynar, D., Demir, I., Goska, R., Kruger, A., 6 (6), 861–867.
Langel, C., Mantilla, R., Niemeier, J., Quintero, F., Seo, B. C. Li, Y., Ye, S. & Bartoli, I.  Semisupervised classification of
& Small, S. J.  Real-time flood forecasting and hurricane damage from postevent aerial imagery using deep
information system for the state of Iowa. Bulletin of the learning. Journal of Applied Remote Sensing 12 (4), 045008.
American Meteorological Society 98 (3), 539–554. Li, L., Jiang, P., Xu, H., Lin, G., Guo, D. & Wu, H.  Water
Kratzert, F., Klotz, D., Brenner, C., Schulz, K. & Herrnegger, M. quality prediction based on recurrent neural network and
 Rainfall–runoff modelling using long short-term memory improved evidence theory: a case study of Qiantang River.
(LSTM) networks. Hydrology and Earth System Sciences China. Environmental Science and Pollution Research
22 (11), 6005–6022. 26 (19), 19879–19896.
Kratzert, F., Klotz, D., Herrnegger, M., Sampson, A. K., Liang, Z., Zou, R., Chen, X., Ren, T., Su, H. & Liu, Y. 
Hochreiter, S. & Nearing, G. S. a Toward improved Simulate the forecast capacity of a complicated water quality
predictions in ungauged basins: exploiting the power of model using the long short-term memory approach. Journal
machine learning. Water Resources Research. of Hydrology 581, 124432.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
33 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Lin, T., Horne, B. G., Tino, P. & Giles, C. L.  Learning long- International Conference on Machine Learning (ICML-10).
term dependencies in NARX recurrent neural networks. pp. 807–814.
IEEE Transactions on Neural Networks 7 (6), 1329–1338. Nam, K. & Wang, F.  An extreme rainfall-induced landslide
Ling, F., Boyd, D., Ge, Y., Foody, G. M., Li, X., Wang, L., Zhang, susceptibility assessment using autoencoder combined with
Y., Shi, L., Shang, C., Li, X. & Du, Y.  Measuring river random forest in Shimane Prefecture, Japan.
wetted width from remotely sensed imagery at the subpixel Geoenvironmental Disasters 7 (1), 6.
scale with a deep convolutional neural network. Water Nam, K., Heo, S., Loy-Benitez, J., Ifaei, P., Lee, S., Kang, S. & Yoo,
Resources Research 55 (7), 5631–5649. C.  An autonomous operational trajectory searching
Liu, P., Wang, J., Sangaiah, A. K., Xie, Y. & Yin, X.  Analysis system for an economic and environmental membrane
and prediction of water quality using LSTM deep neural bioreactor plant using deep reinforcement learning. Water
networks in IoT environment. Sustainability 11 (7), 2058. Science and Technology.
Liu, Y., Piramanayagam, S., Monteiro, S. T. & Saber, E. a Nguyen, D. H. & Bae, D. H.  Correcting mean areal
Semantic segmentation of multisensor remote sensing precipitation forecasts to improve urban flooding predictions
imagery with deep ConvNets and higher-order conditional by using long short-term memory network. Journal of
random fields. Journal of Applied Remote Sensing 13 (1), Hydrology 584, 124710.
016501. Ni, L., Wang, D., Singh, V. P., Wu, J., Wang, Y., Tao, Y. & Zhang, J.
Liu, Y., Qin, H., Zhang, Z., Yao, L., Wang, Y., Li, J., Liu, G. &  Streamflow and rainfall forecasting by two long short-
Zhou, J. b Deriving reservoir operation rule based on term memory-based models. Journal of Hydrology 124296.
Bayesian deep learning method considering multiple O’Neil, C.  Weapons of Math Destruction: How Big Data
uncertainties. Journal of Hydrology 579, 124207. Increases Inequality and Threatens Democracy. Broadway
Ma, L., Liu, Y., Zhang, X., Ye, Y., Yin, G. & Johnson, B. A.  Books.
Deep learning in remote sensing applications: a meta- O’Neil, G. L., Goodall, J. L., Behl, M. & Saby, L.  Deep learning
analysis and review. ISPRS Journal of Photogrammetry and using physically-informed input data for wetland identification.
Remote Sensing 152, 166–177. Environmental Modelling & Software 126, 104665.
Mamandipoor, B., Majd, M., Sheikhalishahi, S., Modena, C. & O’Reilly, T.  Open data and algorithmic regulation. In: Beyond
Osmani, V.  Monitoring and detecting faults in Transparency: Open Data and the Future of Civic Innovation.
wastewater treatment plants using deep learning. pp. 289–300.
Environmental Monitoring and Assessment 192 (2), 148. Pan, B., Hsu, K., AghaKouchak, A. & Sorooshian, S. 
Marçais, J. & de Dreuzy, J. R.  Prospective interest of deep Improving precipitation estimation using convolutional neural
learning for hydrological inference. Groundwater 55 (5), network. Water Resources Research 55 (3), 2301–2321.
688–692. Persello, C., Tolpekin, V. A., Bergado, J. R. & de By, R. A. 
McCuen, R. H.  Modeling Hydrologic Change: Statistical Delineation of agricultural fields in smallholder farms from
Methods. CRC press. satellite images using fully convolutional networks and
Mei, M. J., Maksym, T., Weissling, B. & Singh, H.  Estimating combinatorial grouping. Remote Sensing of Environment
early-winter Antarctic sea ice thickness from deformed ice 231, 111253.
morphology. The Cryosphere 13 (11), 2915–2934. Pollack, J. B.  Recursive distributed representations. Artificial
Meng, X., Zhang, S. & Zang, S.  Lake wetland classification Intelligence 46 (1–2), 77–105.
based on an SVM-CNN composite classifier and high- Poornima, S. & Pushpalatha, M.  Prediction of rainfall using
resolution images using wudalianchi as an example. Journal intensified LSTM based recurrent neural network with
of Coastal Research 93 (sp1), 153–162. weighted linear units. Atmosphere 10 (11), 668.
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Qi, Y., Zhou, Z., Yang, L., Quan, Y. & Miao, Q.  A
Wierstra, D. & Riedmiller, M.  Playing Atari with Deep decomposition-ensemble learning model based on LSTM
Reinforcement Learning. arXiv preprint arXiv:1312.5602. neural network for daily reservoir inflow forecasting. Water
Mo, S., Zabaras, N., Shi, X. & Wu, J. a Deep autoregressive Resources Management 33 (12), 4123–4139.
neural networks for high-dimensional inverse problems in Qin, J., Liang, J., Chen, T., Lei, X. & Kang, A.  Simulating and
groundwater contaminant source identification. Water predicting of hydrological time series based on tensorFlow
Resources Research 55 (5), 3856–3881. deep learning. Polish Journal of Environmental Studies 28 (2).
Mo, S., Zhu, Y., Zabaras, N., Shi, X. & Wu, J. b Deep Rahwan, I.  Society-in-the-loop: programming the algorithmic
convolutional encoder-decoder networks for uncertainty social contract. Ethics and Information Technology 20 (1),
quantification of dynamic multiphase flow in heterogeneous 5–14.
media. Water Resources Research 55 (1), 703–728. Read, J. S., Jia, X., Willard, J., Appling, A. P., Zwart, J. A., Oliver, S.
Moy de Vitry, M., Kramer, S., Wegner, J. D. & Leitão, J. P.  K., Karpatne, A., Hansen, G. J., Hanson, P. C., Watkins, W. &
Scalable flood level trend monitoring with surveillance Steinbach, M.  Process-guided deep learning predictions
cameras using a deep convolutional neural network. of lake water temperature. Water Resources Research.
Hydrology and Earth System Sciences 23 (11), 4621–4634. Reddy, D. S. & Prasad, P. R. C.  Prediction of vegetation
Nair, V. & Hinton, G. E.  Rectified linear units improve dynamics using NDVI time series data and LSTM. Modeling
restricted boltzmann machines. In: Proceedings of the 27th Earth Systems and Environment 4 (1), 409–419.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
34 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Reinsel, D., Gantz, J. & Rydning, J.  The Digitization of the Sermet, Y. & Demir, I.  Virtual and augmented reality
World From Edge to Core. IDC White Paper. applications for environmental science education and
Reisman, D., Schultz, J., Crawford, K. & Whittaker, M.  training. In: New Perspectives on Virtual and Augmented
Algorithmic Impact Assessments: A Practical Framework for Reality: Finding New Ways to Teach in A Transformed
Public Agency Accountability. AI Now Institute, pp. 1–22. Learning Environment. Routledge, London, UK, p. 261–275.
Ren, T., Liu, X., Niu, J., Lei, X. & Zhang, Z.  Real-time water Sermet, Y., Demir, I. & Muste, M. a A serious gaming
level prediction of cascaded channels based on multilayer framework for decision support on hydrological hazards.
perception and recurrent neural network. Journal of Science of The Total Environment 138895.
Hydrology 124783. Sermet, Y., Villanueva, P., Sit, M. A. & Demir, I. b
Rohmat, F. I., Labadie, J. W. & Gates, T. K.  Deep learning for Crowdsourced approaches for stage measurements at
compute-efficient modeling of BMP impacts on stream- ungauged locations using smartphones. Hydrological
aquifer exchange and water law compliance in an irrigated Sciences Journal 65 (5), 813–822.
river basin. Environmental Modelling & Software 122, Shafer, G.  A Mathematical Theory of Evidence, Vol. 42.
104529. Princeton university press.
Ronneberger, O., Fischer, P. & Brox, T.  U-net: convolutional Shen, C.  A transdisciplinary review of deep learning research
networks for biomedical image segmentation. In: and its relevance for water resources scientists. Water
International Conference on Medical Image Computing Resources Research 54 (11), 8558–8593.
and Computer-Assisted Intervention. Springer, Cham, Shen, R., Huang, A., Li, B. & Guo, J.  Construction of a
pp. 234–241. drought monitoring model using deep learning based on
Rumelhart, D. E., Hinton, G. E. & Williams, R. J.  Learning multi-source remote sensing data. International Journal of
Internal Representations by Error Propagation (No. ICS- Applied Earth Observation and Geoinformation 79, 48–57.
8506). California Univ San Diego La Jolla Inst for Cognitive Shi, S. & Xu, G.  Novel performance prediction model of a
Science. biofilm system treating domestic wastewater based on
Rumelhart, D. E., Hinton, G. E. & Williams, R. J.  Learning stacked denoising auto-encoders deep learning network.
representations by back-propagating errors. Nature 323 Chemical Engineering Journal 347, 280–290.
(6088), 533–536. Shin, J., Kim, S. M., Son, Y. B., Kim, K. & Ryu, J. H.  Early
Saba, L., Biswas, M., Kuppili, V., Godia, E. C., Suri, H. S., Edla, prediction of Margalefidinium polykrikoides bloom using a
D. R., Omerzu, T., Laird, J. R., Khanna, N. N., Mavrogeni, S. LSTM neural network model in the South Sea of Korea.
& Protogerou, A.  The present and future of deep Journal of Coastal Research 90 (sp1), 236–242.
learning in radiology. European Journal of Radiology. Simonyan, K. & Zisserman, A.  Very Deep Convolutional
Sahiner, B., Pezeshk, A., Hadjiiski, L. M., Wang, X., Drukker, K., Networks for Large-Scale Image Recognition. arXiv preprint
Cha, K. H., Summers, R. M. & Giger, M. L.  Deep arXiv:1409.1556.
learning in medical imaging and radiation therapy. Medical Sit, M. & Demir, I.  Decentralized Flood Forecasting Using
Physics 46 (1), e1–e36. Deep Neural Networks. arXiv preprint arXiv:1902.02308.
Sankaranarayanan, S., Prabhakar, M., Satish, S., Jain, P., Sit, M., Sermet, Y. & Demir, I. a Optimized watershed delineation
Ramprasad, A. & Krishnan, A.  Flood prediction based library for server-side and client-side web applications. Open
on weather parameters using deep learning. Journal of Water Geospatial Data, Software and Standards 4 (1), 8.
and Climate Change. Sit, M. A., Koylu, C. & Demir, I. b Identifying disaster-related
Santos, J. E., Xu, D., Jo, H., Landry, C. J., Prodanović, M. & Pyrcz, tweets and their semantic, spatial and temporal context using
M. J.  PoreFlow-Net: A 3D convolutional neural network deep learning, natural language processing and spatial
to predict fluid flow through porous media. Advances in analysis: a case study of Hurricane Irma. International
Water Resources 138, 103539. Journal of Digital Earth 12 (11), 1205–1229.
Scher, S. & Messori, G.  Predicting weather forecast Song, A., Kim, Y. & Kim, Y.  Change detection of surface
uncertainty with machine learning. Quarterly Journal of the water in remote sensing images based on fully convolutional
Royal Meteorological Society 144 (717), 2830–2841. network. Journal of Coastal Research 91 (sp1), 426–430.
Sengupta, S., Basak, S., Saikia, P., Paul, S., Tsalavoutis, V., Atiah, Su, A., Li, H., Cui, L. & Chen, Y.  A convection nowcasting
F., Ravi, V. & Peters, A.  A review of deep learning with method based on machine learning. Advances in
special emphasis on architectures, applications and recent Meteorology, 2020.
trends. Knowledge-Based Systems 105596. Sun, A. Y., Scanlon, B. R., Zhang, Z., Walling, D., Bhanja, S. N.,
Sermet, Y. & Demir, I. a An intelligent system on knowledge Mukherjee, A. & Zhong, Z. a Combining physically
generation and communication about flooding. based modeling and deep learning for fusing GRACE satellite
Environmental Modelling & Software 108, 51–60. data: can we learn from mismatch? Water Resources
Sermet, Y. & Demir, I. b Flood action VR: a virtual reality Research 55 (2), 1179–1195.
framework for disaster awareness and emergency response Sun, Z., Zhao, X., Wu, M. & Wang, C. b Extracting urban
training. In Proceedings of the International Conference on impervious surface from worldView-2 and airborne LiDAR
Modeling, Simulation and Visualization Methods (MSV). data using 3D convolutional neural networks. Journal of the
pp. 65–68. Indian Society of Remote Sensing 47 (3), 401–412.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
35 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Sun, Z., Long, D., Yang, W., Li, X. & Pan, Y.  Reconstruction simulations based on machine learning and big data. Journal
of GRACE data on changes in total water storage over the of Hydrology 124875.
global land surface and 60 basins. Water Resources Research Wang, Y., Fang, Z., Hong, H. & Peng, L. d Flood susceptibility
56 (4), e2019WR026250. mapping using convolutional neural network frameworks.
Sundermeyer, M., Schlüter, R. & Ney, H.  LSTM neural Journal of Hydrology 582, 124482.
networks for language modeling. In Thirteenth Annual Weber, L. J., Muste, M., Bradley, A. A., Amado, A. A., Demir, I.,
Conference of the International Speech Communication Drake, C. W., Krajewski, W. F., Loeser, T. J., Politano, M. S.,
Association. Shea, B. R. & Thomas, N. W.  The Iowa watersheds
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., project: Iowa’s prototype for engaging communities and
Erhan, D., Vanhoucke, V. & Rabinovich, A.  Going professionals in watershed hazard mitigation. International
deeper with convolutions. In Proceedings of the IEEE Journal of River Basin Management 16 (3), 315–328.
Conference on Computer Vision and Pattern Recognition. Weber, T., Corotan, A., Hutchinson, B., Kravitz, B. & Link, R. 
pp. 1–9. Deep Learning for Creating Surrogate Models of Precipitation
Tang, G., Long, D., Behrangi, A., Wang, C. & Hong, Y.  in Earth System Models.
Exploring deep neural networks to retrieve rain and snow in Wei, C. C. & Cheng, J. Y.  Nearshore two-step typhoon wind-
high latitudes using multisensor and reanalysis data. Water wave prediction using deep recurrent neural networks.
Resources Research 54 (10), 8253–8278. Journal of Hydroinformatics 22 (2), 346–367.
Tapley, B. D., Bettadpur, S., Watkins, M. & Reigber, C.  The Weyn, J. A., Durran, D. R. & Caruana, R.  Can machines learn
gravity recovery and climate experiment: mission overview to predict weather? using deep learning to predict gridded
and early results. Geophysical Research Letters 31 (9). 500-hPa geopotential height from historical weather data.
Tran, Q. K. & Song, S. K.  Computer vision in precipitation Journal of Advances in Modeling Earth Systems 11 (8),
nowcasting: applying image quality assessment metrics for 2680–2693.
training deep neural networks. Atmosphere 10 (5), 244. Wieland, M., Li, Y. & Martinis, S.  Multi-sensor cloud and
Ukkonen, P. & Mäkelä, A.  Evaluation of machine learning cloud shadow segmentation with a convolutional neural
classifiers for predicting deep convection. Journal of network. Remote Sensing of Environment 230, 111203.
Advances in Modeling Earth Systems 11 (6), 1784–1802. Worland, S. C., Steinschneider, S., Asquith, W., Knight, R. &
Wan, X., Yang, Q., Jiang, P. & Zhong, P. A.  A hybrid model Wieczorek, M.  Prediction and inference of flow duration
for real-time probabilistic flood forecasting using elman curves using multioutput neural networks. Water Resources
neural network with heterogeneity of error distributions. Research 55 (8), 6850–6868.
Water Resources Management 33 (11), 4027–4050. Wu, H., Yang, Q., Liu, J. & Wang, G.  A spatiotemporal deep
Wang, L. P. & Wan, C. R.  Comments on’ The extreme fusion model for merging satellite and gauge precipitation in
learning machine. IEEE Transactions on Neural Networks China. Journal of Hydrology 584, 124664.
19 (8), 1494–1495. Xiang, Z. & Demir, I.  Distributed long-term hourly
Wang, C., Tandeo, P., Mouche, A., Stopa, J. E., Gressani, V., streamflow predictions using deep learning–a case study for
Longepe, N., Vandemark, D., Foster, R. C. & Chapron, B. state of Iowa. Environmental Modeling & Software 104761.
a Classification of the global Sentinel-1 SAR vignettes for Xiang, Z., Yan, J. & Demir, I.  A rainfall-runoff model with
ocean surface process studies. Remote Sensing of LSTM-based sequence-to-sequence learning. Water
Environment 234, 111457. Resources Research 56 (1).
Wang, J. H., Lin, G. F., Chang, M. J., Huang, I. H. & Chen, Y. R. Xiao, C., Chen, N., Hu, C., Wang, K., Xu, Z., Cai, Y., Xu, L., Chen, Z. &
b Real-time water-level forecasting using dilated causal Gong, J. a A spatiotemporal deep learning model for sea
convolutional neural networks. Water Resources surface temperature field prediction using time-series satellite
Management 33 (11), 3759–3780. data. Environmental Modelling & Software 120, 104502.
Wang, P., Yao, J., Wang, G., Hao, F., Shrestha, S., Xue, B., Xie, G. Xiao, H., Zhang, F., He, Q., Liu, P., Yan, F., Miao, L. & Yang, Z.
& Peng, Y. c Exploring the application of artificial b Classification of ice crystal habits observed from
intelligence technology for identification of water pollution airborne Cloud Particle Imager by deep transfer learning.
characteristics and tracing the source of water quality Earth and Space Science 6 (10), 1877–1886.
pollutants. Science of The Total Environment 693, 133440. Xu, Z., Ying, Z., Li, Y., He, B. & Chen, Y.  Pressure prediction
Wang, J., Yuan, Q., Shen, H., Liu, T., Li, T., Yue, L., Shi, X. & and abnormal working conditions detection of water supply
Zhang, L. a Estimating snow depth by combining network based on LSTM. Water Supply.
satellite data and ground-based observations over Alaska: a Yang, S., Yang, D., Chen, J. & Zhao, B. a Real-time reservoir
deep learning approach. Journal of Hydrology 585, 124828. operation using recurrent neural networks and inflow
Wang, N., Zhang, D., Chang, H. & Li, H. b Deep learning of forecast from a distributed hydrological model. Journal of
subsurface flow via theory-guided neural network. Journal of Hydrology 579, 124229.
Hydrology 584, 124700. Yang, T., Sun, F., Gentine, P., Liu, W., Wang, H., Yin, J., Du, M. &
Wang, Q., Huang, J., Liu, R., Men, C., Guo, L., Miao, Y., Jiao, L., Liu, C. b Evaluation and machine learning improvement
Wang, Y., Shoaib, M. & Xia, X. c Sequence-based of global hydrological model-based flood simulations.
statistical downscaling and its application to hydrologic Environmental Research Letters 14 (11), 114027.

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user
 Uncorrected Proof
36 M. Sit et al. | A comprehensive review of deep learning applications in hydrology Water Science & Technology | in press | 2020

Yan, Q., Ji, F., Miao, K., Wu, Q., Xia, Y. & Li, T.  support vector regression, deep learning algorithm. Journal of
Convolutional residual-attention: a deep learning approach Hydrology 565, 720–736.
for precipitation nowcasting. Advances in Meteorology, 2020. Zhang, J., Zhu, Y., Zhang, X., Ye, M. & Yang, J. f Developing a
Yang, F., Feng, T., Xu, G. & Chen, Y.  Applied method for long short-term memory (LSTM) based model for predicting
water-body segmentation based on mask R-CNN. Journal of water table depth in agricultural areas. Journal of Hydrology
Applied Remote Sensing 14 (1), 014502. 561, 918–929.
Yim, I., Shin, J., Lee, H., Park, S., Nam, G., Kang, T., Cho, K. H. & Zhang, X., Zhang, Q., Zhang, G., Nie, Z., Gui, Z. & Que, H.  g
Cha, Y.  Deep learning-based retrieval of cyanobacteria A novel hybrid data-driven model for daily land surface
pigment in inland water for in-situ and airborne temperature forecasting using long short-term memory neural
hyperspectral data. Ecological Indicators 110, 105879. network based on ensemble empirical mode decomposition.
Yin, W., Kann, K., Yu, M. & Schütze, H.  Comparative Study International Journal of Environmental Research and Public
of cnn and rnn for Natural Language Processing. arXiv Health 15 (5), 1032.
preprint arXiv:1702.01923. Zhang, C., Li, G., Du, S., Tan, W. & Gao, F. a Three-
Yu, Z., Yang, K., Luo, Y. & Shang, C.  Spatial-temporal process dimensional densely connected convolutional network for
simulation and prediction of chlorophyll-a concentration in hyperspectral remote sensing image classification. Journal of
Dianchi Lake based on wavelet analysis and long-short term Applied Remote Sensing 13 (1), 016519.
memory network. Journal of Hydrology 582, 124488. Zhang, Z., Iwasaki, A., Xu, G. & Song, J. b Cloud detection on
Yuan, X., Chen, C., Lei, X., Yuan, Y. & Adnan, R. M.  Monthly small satellites based on lightweight U-net and image
runoff forecasting based on LSTM–ALO model. Stochastic compression. Journal of Applied Remote Sensing 13 (2),
Environmental Research and Risk Assessment 32 (8), 026502.
2199–2212. Zhang, C. J., Zeng, J., Wang, H. Y., Ma, L. M. & Chu, H. 
Yurtsever, M. & Yurtsever, U.  Use of a convolutional neural Correction model for rainfall forecasts using the LSTM with
network for the classification of microbeads in urban multiple meteorological factors. Meteorological Applications
wastewater. Chemosphere 216, 271–280. 27 (1), e1852.
Zeng, W., Zhang, D., Fang, Y., Wu, J. & Huang, J.  Zhou, Y. T. & Chellappa, R.  Computation of optical flow
Comparison of partial least square regression, support vector using a neural network. In IEEE International Conference on
machine, and deep-learning techniques for estimating soil Neural Networks. Vol. 1998, pp. 71–78.
salinity from hyperspectral data. Journal of Applied Remote Zhou, X., Tang, Z., Xu, W., Meng, F., Chu, X., Xin, K. & Fu, G.
Sensing 12 (2), 022204.  Deep learning identifies accurate burst locations in
Zhang, D., Hølland, E. S., Lindholm, G. & Ratnaweera, H. a water distribution networks. Water Research 166, 115058.
Hydraulic modeling and deep learning based flow forecasting Zhou, Z., Shi, L. & Zha, Y.  Seeing macro-dispersivity from
for optimizing inter catchment wastewater transfer. Journal hydraulic conductivity field with convolutional neural
of Hydrology 567, 792–802. network. Advances in Water Resources 103545.
Zhang, D., Lindholm, G. & Ratnaweera, H. b Use long short- Zhu, S., Hrnjica, B., Ptak, M., Choiński, A. & Sivakumar, B. a
term memory to enhance internet of things for combined Forecasting of water level in multiple temperate lakes using
sewer overflow monitoring. Journal of Hydrology 556, machine learning models. Journal of Hydrology 124819.
409–418. Zhu, S., Luo, X., Yuan, X. & Xu, Z. b An improved long short-
Zhang, D., Martinez, N., Lindholm, G. & Ratnaweera, H. c term memory network for streamflow forecasting in the
Manage sewer in-line storage control using hydraulic model upper Yangtze River. Stochastic Environmental Research and
and recurrent neural network. Water Resources Management Risk Assessment 1–17.
32 (6), 2079–2098. Zou, Q., Xiong, Q., Li, Q., Yi, H., Yu, Y. & Wu, C.  A water
Zhang, C., Sargent, I., Pan, X., Li, H., Gardiner, A., Hare, J. & quality prediction method based on the multi-time scale
Atkinson, P. M. d An object-based convolutional neural bidirectional long short-term memory network.
network (OCNN) for urban land use classification. Remote Environmental Science and Pollution Research 1–12.
Sensing of Environment 216, 57–70. Zuo, G., Luo, J., Wang, N., Lian, Y. & He, X.  Decomposition
Zhang, D., Lin, J., Peng, Q., Wang, D., Yang, T., Sorooshian, S., ensemble model based on variational mode decomposition
Liu, X. & Zhuang, J. e Modeling and simulating of and long short-term memory for streamflow forecasting.
reservoir operation using the artificial neural network, Journal of Hydrology 124776.

First received 16 June 2020; accepted in revised form 22 July 2020. Available online 5 August 2020

Downloaded from https://iwaponline.com/wst/article-pdf/doi/10.2166/wst.2020.369/721217/wst2020369.pdf

by Western Sydney University user