998 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL.

16, 2023

Susceptibility-Guided Landslide Detection Using

Fully Convolutional Neural Network
Yangyang Chen , Dongping Ming , Member, IEEE, Junchuan Yu, Lu Xu, Yanni Ma, Yan Li,
Xiao Ling , and Yueqin Zhu

Abstract—Automatic landslide detection based on very high I. INTRODUCTION

spatial resolution remote sensing images is crucial for disaster pre-
EASONAL torrential rainfall triggers vast landslides on the
vention and mitigation applications. With the rapid development
of deep-learning techniques, state-of-the-art semantic segmenta-
tion methods based on fully convolutional network (FCNN) have
S southeast coast of China every year, especially in moun-
tainous terrains [1]. It is worth noting that with the rapid growth
achieved outstanding performance in the landslide detection task. of urbanization and city expansion, the mountainous area once
However, most of the existing articles only utilize visual features.
Even if the advanced FCNN models are applied, there is still a
inhospitable has gradually changed into urban infrastructures.
certain amount of falsely detected and miss detected landslides. In Therefore, landslide occurrence has posed an increasing threat
this article, we innovatively introduce landslide susceptibility as to lives and property on the southeast coast of China in recent
prior knowledge and propose an innovative susceptibility-guided years. Taking Hong Kong Special Administration Region as an
landslide detection method based on FCNN (SG-FCNN) to detect example, two fatalities were reported, and more than a hundred
landslides from single temporal images. In addition, an unsuper-
vised change detection method based on the mean changing mag-
municipal roads were damaged or closed as a result of a landslide
nitude of objects (MCMO) is further proposed and integrated with triggered by the rainstorm on 7th June 2008 [2]. To alleviate the
the SG-FCNN to detect newly occurred landslides from bitemporal negative impact of landslides and landslide-induced geohazards,
images. The effectiveness of the proposed SG-FCNN and MCMO industry and academia need to conduct in-depth research on
has been tested in Lantau Island, Hong Kong. The experimental landslide monitoring, susceptibility analysis, and early warning
results show that the SG-FCNN can significantly reduce the amount
of falsely detected and miss detected landslides compared with the
[3], [4].
FCNN. It can conclude that applying landslide susceptibility as Landslide inventories are the basis for carrying out the re-
prior knowledge is much more effective than using visual features search mentioned above [4], [5] and can be obtained via landslide
only, which introduces a new methodology of landslide detection detection [6]. Therefore, it is vital to detect landslides precisely,
and lifts the detection performance to a new level. quickly, and automatically [7]. In the early stages, landslide
Index Terms—Convolutional neural network (CNN), landslide detection mainly relied on geological field survey [8]. Field
detection, landslide susceptibility mapping, Lantau Island, remote survey can get highly reliable landslide information, but it is
sensing. labor intensive [9] and can only be applied on a site-specific
or local scale [4]. With the rapid development of satellite and
Manuscript received 30 August 2022; revised 8 December 2022; accepted 27 sensor techniques since the 1970s, remote sensing products
December 2022. Date of publication 30 December 2022; date of current version have been widely applied in the geoscience field [10], making
9 January 2023. This work was supported in part by the National Key Research it possible to detect landslides on a regional scale. Initially,
and Development Program of China under Grant 2021YFC3000400, in part
by the Fundamental Research Funds for the Central Universities, in part by the remote-sensing-based landslide detection primarily relied on
Advance Research Project of Civil Space Technology, and in part by the National visual interpretations performed by geological experts [11]. In
Natural Science Foundation of China under Grant 41872253. (Corresponding the 21st century, the launch of Google Earth and the emergence
author: Dongping Ming.)
Yangyang Chen and Junchuan Yu are with the China Aero Geophysical Survey of open-source data have made it much easier to acquire remote
and Remote Sensing Center for Natural Resources, Beijing 100083, China sensing products with different sources and modalities [12],
(e-mail: jimmyxiyangyang@hotmail.com; yujunchuan@mail.cgs.gov.cn). [13]. Although the promotion of remote sensing products im-
Dongping Ming is with the School of Information Engineering, China Uni-
versity of Geosciences, Beijing 100083, China, and also with the Polytechnic proves the efficiency of landslide detection, visual interpretation
Center for Natural Resources Big-data, Ministry of Natural Resources of China, still requires a considerable amount of human consumption [6],
Beijing 100036, China (e-mail: mingdp@cugb.edu.cn). which leads to undesirable detection efficiency on a regional
Lu Xu, Yan Li, and Xiao Ling are with the School of Information En-
gineering, China University of Geosciences, Beijing 100083, China (e-mail: scale [14]. Besides, visual interpretation based on expertise and
xlirs@cugb.edu.cn; liyanturbo@163.com; lingx0527@163.com). experience is subjective [11], [15]. Hence, visual interpretation
Yanni Ma is with the China Aero Geophysical Survey and Remote Sensing failed to satisfy the demand for rapid landslide detection on a
Center for Natural Resources, Beijing 100083, China, and also with the School
of Information Engineering, China University of Geosciences, Beijing 100083, regional scale [4].
China (e-mail: mayanni@mail.cgs.gov.cn). With advances in computer hardware and pattern recognition
Yueqin Zhu is with the National Institute of Natural Hazards, Ministry techniques, the accuracy of computer vision algorithms
of Emergency Management, Beijing 100085, China (e-mail: yueqinzhu@
163.com). gradually approaches visual interpretation in various domains,
Digital Object Identifier 10.1109/JSTARS.2022.3233043 which also benefits landslide detection. Remote-sensing-based

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
CHEN et al.: SUSCEPTIBILITY-GUIDED LANDSLIDE DETECTION USING FULLY CONVOLUTIONAL NEURAL NETWORK 999

landslide detection is now moving from visual interpretation PBCNN effectively and has drawn the attention of an increasing
toward automatic interpretation [16], among which machine number of scholars. Meena et al. [41] compared the detection
learning (ML)-based methods are the most representative. By performance of U-Net, SVM, K-nearest neighbor, and RF in the
extracting and utilizing low-level features of remote sensing Rasuwa district, Nepal. The U-Net achieved higher accuracy
images, various traditional ML models, such as maximum than other traditional ML models. Prakash et al. [42] modified
likelihood [17], support vector machine (SVM) [18], [19], and U-Net using ResNet34 as the backbone and achieved better
random forest (RF) [20], have raised the efficiency of landslide performance than pixel-based and object-based ML methods
detection with acceptable accuracy. The spatial resolution in Douglas County, USA. Liu et al. [43] modified U-Net by
of remote sensing images has significantly increased over adding a residual learning unit and used it to improve the
the last few decades. However, very high spatial resolution accuracy of postearthquake landslide detection in Jiuzhaigou,
(VHR) remote sensing images suffer severe interclass similarity China. Braganolo et al. [40] utilized U-Net and the compound
[21], [22]. For example, muddy tracks and dry riverbanks are loss function to detect landslide scars from Landsat 8 images,
indistinguishable from landslides because they share similar and the result outperformed RF and the PBCNN. Ghorbanzadeh
spectral features. Interclass similarity poses a conundrum for et al. [44] proposed an object-based image analysis (OBIA)-
traditional ML algorithms, and it is challenging to obtain ideal based ResU-Net to detect landslides in east Iburi, Japan, which
landslide detection results by utilizing only low-level features. achieved 22% higher intersection over union (IoU) than the
Mining the high-level features of VHR remote sensing images original ResU-Net. Liu et al. [45] proposed an improved Mask
is the key to improving detection accuracy and reliability. R-CNN models for landslide detection in Jiuzhaigou, China, and
After entering the new century, numerous powerful deep- achieved satisfactory accuracy with precision, recall, and overall
learning (DL) models have been launched and put to scientific accuracy (OA) of 95.8%, 93.1%, and 94.7%, respectively.
use. As the most mature and widely applied DL model, the To date, scholars have been trying to apply advanced DL
convolutional neural network (CNN) has exceptional advantages models to landslide detection, and the results show that DL
in high-level feature extraction and representation [23], which models are superior to traditional ML models, especially the
makes up for the defects of traditional ML models [24]. CNN FCNN. However, since landslides are indistinguishable from
has now been widely used in the geoscience domain, such as some land cover features, false and miss detection still exist.
scene classification [25], land-cover classification [26], [27], One reason is the lack of datasets containing vast landslide
change detection [28], [29], and ground target detection [30]. samples with all kinds of features combined with the negative
In recent years, many scholars have applied CNN in landslide- samples. Although adopting a data augmentation strategy can
related domains, such as landslide susceptibility mapping [31], expand small datasets to a certain extent, detection results are
[32], landslide deformation prediction [33], [34], and especially still less than satisfactory.
landslide detection. There are mainly two types of CNN-based Introducing auxiliary information is a promising strategy
landslide detection models: Patch-based CNN (PBCNN) and to improve landslides’ detection accuracy further. Some re-
fully CNN (FCNN) [35]. searchers [16], [42], [46], [47], [48], [49], [50] fed remote
Yu et al. [36] first introduced the PBCNN and region growth sensing images and auxiliary information, such as elevation,
algorithm to detect landslides. Ghorbanzadeh et al. [37] ap- aspect, land cover, and normalized difference vegetation index
plied PBCNN to detect landslides in Nepal from Rapid Eye (NDVI) into CNN. It allows CNN to explore richer high-level
images and has proved that the PBCNN outperforms traditional features, especially the connection between landslide occurrence
ML methods, including SVM, RF, and multilayer perceptron. and surrounding environmental conditions, achieving higher
Sameen and Pradhan [38] utilized the PBCNN with advanced detection accuracy. However, the auxiliary information is in-
residual blocks to detect landslides in Malaysia’s Cameron high- dispensable for training and testing regions. Besides, accurate
lands based on spectral features and topographical information geometric registration of different auxiliary information layers
derived from remote sensing products. Ji et al. [39] introduced and correlated remote sensing images are also necessary. Mean-
the three-dimensional attention-boosted PBCNN model to de- while, the structure of the classical CNN needs to be redesigned
tect landslides in Zhijin County, China, and achieved higher carefully. Applying auxiliary information to the postprocessing
detection accuracy than other recent attention-based models. of landslide detection results is another solution. For example,
Although the PBCNN has made significant progress in the Shi et al. [51] used a binary mask operation to optimize landslide
automatic intelligent detection of landslides, defects still exist. detection results based on road vectors, building vectors, and
The principle of the PBCNN is to classify each pixel one by slope maps. The results showed that the precision and F1-score
one, whose detection results will suffer from salt–pepper noise. were substantially improved. However, the selection of auxiliary
In addition, for each pixel, a square image patch is clipped information and the determination of corresponding thresholds
and fed into CNN for classification, which leads to extremely are often subjective. In addition, applying a specific type of
low efficiency and a heavy computational burden [26], thus auxiliary information is unreasonable. For instance, a binary
limiting the efficiency of the PBCNN for detecting landslides mask operation may mistakenly remove detected landslides that
on a regional scale. crossed the road.
The FCNN can generate pixelwise detection results at once The landslide detection methods reviewed above are based
[40], not predicting only the central pixel of the input image on single temporal images and are feasible for generating the
patch. Therefore, the FCNN can avoid the shortage of the initial version of the landslide inventory. In actual applications,
1000 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 16, 2023

Fig. 1. Location of Lantau Island.

detecting newly occurred landslides is equally essential [6]. With bitemporal remote sensing images. It is also the first time that the
the help of bitemporal (postlandslide and prelandslide) images, unsupervised change detection method has been integrated with
the boundaries and occurrence time of newly occurred landslides the FCNN to detect newly occurred landslides from bitemporal
can be obtained precisely [52]. Some researchers have recently images.
applied the FCNN to detect newly occurred landslides. Stacking This article is organized as follows. Section II introduces the
the postlandslide and prelandslide images and feeding it into the study area. Section III presents an in-depth description of the
FCNN is the most widely applied strategy. However, the spectral proposed methods. Section IV shows the experimental details
difference of bitemporal images extracted by shallow layers may and results. Sections V and VI provide the discussion and
not represent semantic change and mistakenly propagate to deep conclusion, respectively.
layers, leading to false detection [53]. Shi et al. [51] first used
the FCNN to detect landslides from postlandslide and preland-
slide images, respectively. The newly occurred landslides were II. STUDY AREA
detected by utilizing the change detection technique based In this article, Lantau Island, Hong Kong (shown in Fig. 1),
on OBIA, which outperformed unsupervised change detection which is situated on the southeast coast of China, is selected
methods based on the changing magnitude image (CMI) and as the study area. As the largest outlying island in Hong Kong
all CNN-based rivals. Nevertheless, for all supervised methods, [56], Lantau Island is dominated by mountainous terrain and
including the FCNN, detecting newly occurred landslides from composed of volcanic rock and granite.
bitemporal images require a much larger amount of labor on There are three reasons for selecting Lantau Island as the
sample labeling than detecting landslides from single-temporal study area. First, Lantau Island has a subtropical monsoon
images, leading to lower processing efficiency. climate with 2200 mm of average annual rainfall and frequent
To address the abovementioned issues, we propose a novel torrential rainfall, which triggers a vast number of landslides
susceptibility-guided landslide detection method based on every year. Especially in June 2008, a record-breaking rainstorm
FCNN (SG-FCNN). Specifically, the landslide susceptibility triggered 2610 landslides (shown in Fig. 2), causing havoc and
map generated by the state-of-the-art CNN model is intro- two fatalities on Lantau Island [57]. The types of natural terrain
duced as prior knowledge, which guides the FCNN to detect landslides on Lantau Island include debris flows, rockfalls, and
landslides from single-temporal remote sensing images. The slumps, among which debris flows appear most frequently [58].
landslide susceptibility map can provide the spatial probability Second, with the expansion of urban areas, an increasing number
of landslide occurrence in the near future [54], [55], which is of infrastructures and landmarks have been built on Lantau
obtained based on the analysis of landslide-related auxiliary Island over the past decades. Most of them are densely pop-
information (also known as landslide predisposing factors). To ulated, such as Chek Lap Kok International Airport, Hong Kong
the best of our knowledge, this is the first time that landslide sus- Disneyland Resort, Tung Chung New Town, Discovery Bay,
ceptibility has been introduced as prior knowledge to guide the and Ngong Ping. Therefore, landslide prevention and mitigation
detection of landslides. Additionally, an unsupervised change on Lantau Island is crucial for Hong Kong authorities. Third,
detection method based on the mean changing magnitude of to carry out landslide-related applications and research, the
objects (MCMO) is proposed and jointly used with the proposed Civil Engineering and Development Department of Hong Kong
SG-FCNN, which expands the application field of the proposed established the Enhanced Natural Terrain Landslide Inventory
SG-FCNN to the detection of newly occurred landslides from (ENTLI) based on annually updated digital aerial photos (DAPs)
CHEN et al.: SUSCEPTIBILITY-GUIDED LANDSLIDE DETECTION USING FULLY CONVOLUTIONAL NEURAL NETWORK 1001

detected by the proposed MobileU-Net under the guidance of

landslide susceptibility. The boundaries of detected landslides
were then optimized by the dense conditional random field
(Dense CRF). Finally, to detect newly occurred landslides in
2008, bitemporal DAPs and the proposed MCMO were used
to identify and remove historical landslides from the detection
results of all landslides.

A. Data Collection and Preprocessing

1) Digital Aerial Photos: As shown in Fig. 4, a subset region
of central Lantau Island was selected as the study area. Two
testing regions are highlighted with red and green boundaries,
and the other remaining areas are training regions. Six scenes of
DAPs taken by the Geotechnical Engineering Office were pur-
chased and applied, four of which (Photo numbers: CW82126,
Fig. 2. Landslides triggered by the rainstorm in June 2008 [2]. (a) Nam Chung
Tsuen, Tai O. (b) North Lantau highway. (c) Keung Shan Kwum Yum Temple.
CW82130, CW82134, and CW82147) covered the whole study
(d) Sham Wat road. area were acquired in November 2008. The other two scenes
(Photo numbers: CW79890 and CW79894) covered two testing
regions were acquired in November 2007. All applied DAPs are
via visual interpretation by experts [59]. The latest version of digitized RGB aerial photographic films with a spatial resolution
the ENTLI contains the crowns and trails of 5340 landslides of 0.5 m, which were acquired by a Zeiss RMK TOP 15 aerial
covered from 1974 to 2015 on Lantau Island. Related DAPs survey camera at a flying height of 8000 ft.
are purchasable on the Hong Kong Map Service 2.0 website The data preprocessing of the DAPs was carried out in ENVI
(https://www.hkmapservice.gov.hk). Given the above, Lantau 5.1, including rational polynomial coefficient (RPC) generation,
Island is an ideal study area for landslide detection. orthorectification, mosaicking, bitemporal image registration,
and relative radiometric correction. First, 208 ground control
points for exterior orientation were selected using Google Earth
III. METHODOLOGY
Pro. Based on which RPCs were generated as a reference
This article proposed the SG-FCNN to detect landslides on to the calibration certificate. After that, orthorectification was
Lantau Island in 2008. Fig. 3 presents the detailed workflow. The conducted based on prepared RPCs and digital terrain model
general outline is divided into landslide susceptibility mapping with a spatial resolution of 5 m. Then, the seamless mosaic
and landslide detection. technique was adopted to stitch the four scenes of the 2008 DAPs
For landslide susceptibility mapping, a series of landslide- into a single image. Finally, the automatic histogram matching
related predisposing factors were first collected. Afterward, a technique was applied for relative radiometric correcting of the
state-of-the-art DL-based landslide susceptibility model, the en- bitemporal DAPs.
semble model based on channel-expanded pretrained CNN and 2) Training Dataset: The proposed SG-FCNN is based on
RF (CPCNN-RF) [31], was trained based on the predisposing semantic segmentation, and a pixelwise labeled reference map
factor layers and the ENTLI (from 1994 to 2007). Finally, the is required. The ENTLI only records landslide crowns and trails
trained CPCNN-RF was applied to map the landslide suscep- instead of the boundaries and cannot be directly used as a label
tibility of Lantau Island. The generated landslide susceptibility for semantic segmentation. Therefore, landslide boundaries in
map can be used to predict landslide occurrence after 2007. The the training regions were visually delineated as polygons by
methodology of the CPCNN-RF and the landslide susceptibility reference to the DAPs (2008) and the ENTLI (from 1994 to
mapping process in Lantau Island have been detailed in [31]. 2008). In the training regions, 672 landslide polygons were
Therefore, this article will not go into detail. drawn and rasterized into binary label maps.
For landslide detection, there were two missions in this article: To make up the training dataset, DAPs and their corresponding
1) detecting all landslides from single-temporal aerial images binary label maps were clipped into square patches with a size
acquired in 2008, which includes historical landslides before of 416 × 416. In most previous articles, the grid-clipping
2008 and newly occurred landslides in 2008. 2) Detecting strategy is performed to generate square patches. However,
newly occurred landslides in 2008 from bitemporal aerial images considering the arear of training regions are relatively small and
acquired in 2007 and 2008. The related data were collected the complexity of the proposed SG-FCNN, the grid clipping
first, including the ENTLI (from 1994 to 2008) and the DAPs strategy can only generate 327 sample patches, which is not
collected in 2007 and 2008. Then, landslide polygons were enough for the SG-FCNN training. Hence, an overlapping
delineated by taking the record in the ENTLI (from 1994 to clipping strategy was applied; each clipped patch shared a 75%
2008) as referees, and an augmented training dataset along with overlapping area with the horizontally and vertically adjacent
DAPs (2008) was produced. After that, all landslides (including patches. In total, 5139 sample patches were generated in the
historical and newly occurred landslides) in DAPs (2008) were training regions. To further improve the generalization ability of
1002 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 16, 2023

Fig. 3. Flowchart of the proposed SG-FCNN.

B. Landslide Detection Using FCNN

1) MobileU-Net: The convolutional encoder–decoder archi-
tecture is utilized in most of the existing FCNN models. The
high-level feature map of the input image is extracted in the
encoder part using convolutional feature extractors, and the
label of each pixel in the input image is predicted in the de-
coder part. U-Net [60] is one of the most successful convo-
lutional encoder–decoder architectures, and has been applied
in various tasks, including landslide detection. Compared to
the traditional encoder–decoder architecture, U-Net applies the
skip connection technique, which involves feature maps with
different abstract levels in the decoder process. The skip con-
nection effectively boosts semantic segmentation performance
and substantially reduces the feature loss of the convolution
operation in deep layers. The performance of the convolutional
Fig. 4. Scope of study area and testing regions. encoder–decoder-based models partly depends on the encoder
structure. But the encoder of the original U-Net is the VGG-16,
which is heavy and inefficient.
the SG-FCNN, data augmentation algorithms including flipping Inspired by the characteristics of U-Net, we modified the
(horizontal/vertical), rotation (90°/180°/270°), averaging blur, original U-Net and proposed the MobileU-Net (Fig. 5) to achieve
Gaussian blur, bilateral blur, and adding random noise were the balance between performance and efficiency in the landslide
adopted. Since landslides covered less than 10% of the training detection task. The lightweight MobileNet-V2 [61] was applied
regions, data augmentation was only applied to square sample as the backbone encoder for feature map extraction in the U-Net.
patches with landslide labels to balance the number of positive Compared to the VGG-16 applied by the original U-Net, the
and negative pixels. In the end, 30 683 sample patches were MobileNet-V2 introduced a depthwise separable convolution
generated and were further divided into training samples and block for feature map extraction, which can effectively save
validation samples at a proportion of 4:1. computation resources. The depthwise separable convolution
CHEN et al.: SUSCEPTIBILITY-GUIDED LANDSLIDE DETECTION USING FULLY CONVOLUTIONAL NEURAL NETWORK 1003

Fig. 5. Structure of mobile U-Net.

separable convolution blocks were concatenated with the first,

second, and third unsampled feature maps, respectively. In
addition, considering that deconvolution operation has more
trainable parameters compared with the combination of unpool-
ing and traditional convolutional operation, the latter can also
effectively avoid the checkerboard effect [62] and was applied
for feature map upsampling. Finally, feature maps extracted by
the last convolution layer in the decoder part were fed into the
Softmax classifier to generate the pixelwise probability map of
landslide detection.
Considering that the augmented dataset produced in Sec-
tion III-A-2 was still too small to train MobileU-Net from
scratch fully, a fine-tuning strategy was applied. The convolution
parameters of MobileNet-V2 pretrained by the ImageNet image
classification dataset were reloaded before the network training.
Fig. 6. Schematic of traditional convolution and depthwise convolution. (a) Moreover, applying data augmentation to square sample patches
Traditional convolution. (b) Depthwise convolution. with landslide labels enlarged the proportion of positive pixels,
the percentage of negative pixels remained high. It may have led
to MobileU-Net becoming more likely to misclassify landslide
block was composed of a depthwise convolution and a 1 × 1 pixels into non-landslide pixels.
original convolution (also called pointwise convolution). As The Dice loss [63] and the weighted binary cross-entropy
shown in Fig. 6, by applying a single filter to each input channel, (BCE) loss [64] can both effectively avoid negative impacts
the depthwise convolution can extract high-level features in caused by imbalanced samples in the training process. Signif-
the spatial dimension and depth dimension. The subsequent icantly, the Dice loss can optimize the F1 -score of detected
pointwise convolution created a linear combination of depthwise landslides directly. Thus, the compound loss function of the Dice
convolution extracted features. loss and the weighted BCE loss was applied during the training
In the encoder part of MobileU-Net, a 3 × 3 original con- process of the MobileU-Net, which is defined as follows:
volution and a total of 11 depthwise separable convolution
blocks (with 3 × 3 deepwise convolution) were used for hierar-
Compound Loss = Weighted BCE Loss + Dice Loss
chal high-level feature extraction. During the decoder process,
feature maps extracted by the first, third, and fifth depthwise = − [w (1 − yi,j ) log (ti,j )
1004 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 16, 2023

Fig. 7. Schematic of overlapping mosaic strategy.

+ (1 − w) yi,j log ti,j ] + 1 methods applied to the training data. For each augmentation
 method, a probability map of landslide detection was obtained
2 ti,j yi,j + 
−  (1) using FCNN and the overlapping mosaic detection strategy.
ti,j + yi,j +  Then, all probability maps based on different augmentation
for the pixel on the ith row and the jth column of each sample, methods were averaged

and yi,j and ti,j represent the predicted label and the ground truth PPi
AP (Pc ) = i (2)
label, respectively. w stands for the ratio of landslide pixels to the i
total in the training dataset.  is a constant with a Keras default where Pc stands for the cth pixel in testing region. P Pi is the
value of 10−7 . prediction probability of the ith augmentation method. AP(Pc )
2) Overlapping Mosaic Detection Strategy: After the train- denotes the averaged prediction probability of Pc . If AP(Pc )
ing process, MobileU-Net is prepared for landslide detection. exceeds 20%, the pixel Pc would be kept as landslide candidates,
Limited by the network structure, the input data size should while others were considered fuzzy boundaries and were dis-
have been restricted to 416 × 416 × 3. Therefore, images of the carded. Finally, the generated candidate pixels were converted
testing regions cannot be fed into the MobileU-Net directly. Ap- into landslide candidate objects based on the eight-neighborhood
plying a grid-clipping strategy was a viable solution. However, pixel-object conversion criteria [51], [65].
the reliability of the landslide detection result near the boundary 2) Landslide Detection Under the Guidance of Susceptibil-
of the clipped patch was relatively poor. The problem was severe ity: Even though the proposed FCNN has applied advanced DL
for some landslides that were clipped and divided into multiple techniques, false detection of other land cover features, such as
patches. An overlapping detection mosaic strategy was applied muddy roads and buildings that could be misclassified as land-
(as shown in Fig. 7) to solve the above issues. slides, remained inevitable. Hence, additional postprocessing is
First, images of testing regions were clipped into several necessary to remove falsely detected landslides.
overlapped square image patches with a size of 416 × 416 × 3, Landslide susceptibility maps can portray the spatial proba-
which was the same as the overlapping clipping strategy applied bility of landslide occurrence in the near future [54], [55], which
in the training dataset generation. Then, the clipped patches is obtained based on the analysis of landslide-related auxiliary
were fed into the MobileU-Net for pixelwise landslide detection. information (also known as landslide predisposing factors) [66].
Next, the central region of the predicted patch was designated as Landslide susceptibility mapping methods have been reviewed
the target area and clipped. Finally, a sliding-window strategy comprehensively by some scholars [5], [54], [67] and thus are not
was applied to mosaic the detection result of predicted patches covered in this article. Recent articles have shown that methods
into seamless detection results. By utilizing the overlapping based on CNN also perform the best in landslide susceptibility
mosaic detection strategy, only the highly reliable central region mapping [31], [68], [69], [70]. If the CNN-predicted landslide
of the predicted patch was retained. susceptibility can be used as prior knowledge in the detection
process, it is promising to further improve the detection perforce
C. Landslide Detection Using SG-FCNN of FCNN.
Thus, the CNN-predicted landslide susceptibility was intro-
To further boost the detection performance of the FCNN, this
duced as prior knowledge to remove falsely detected landslides
section innovatively proposed the SG-FCNN by applying the
as follows:
probability average test time augmentation PA-TTA and the SG 
strategy. The MCMO change detection strategy is also proposed [LS (Pb ) × LDP (Pb )]
SLDP (Oa ) = , Pb ∈ Oa . (3)
to detect newly occurred landslides using SG-FCNN. N
1) Probability Average Test Time Augmentation (PA-TTA): For the ath landslide candidate object Oa , the Pb and N
To reduce the miss detection rate of the FCNN, the PA-TTA was stand for the bth pixel and the total number of pixels in Oa ,
first proposed. The overlapping clipped image patches in the respectively. For pixel Pb , LDP(Pb ) represents the landslide
testing region were augmented using the same augmentation detection probability of Pb and LS(Pb ) denotes the probability
CHEN et al.: SUSCEPTIBILITY-GUIDED LANDSLIDE DETECTION USING FULLY CONVOLUTIONAL NEURAL NETWORK 1005

value of CNN-predicted landslide susceptibility. SLDP(Oa ) After the Dense CRF optimized the boundaries of the detected
represents the SG landslide detection probability of Oa , and landslides, two widely used morphological optimization algo-
is obtained by averaging LS(Pb ) × LDP(Pb ). The threshold of rithms, namely, area opening and hole filling [73], [74], were
susceptibility guidance, TSLDP , was set as 20% based on the implemented to improve the detection accuracy further. The area
division of five equal intervals, which indicated a very low opening was applied to remove those detected landslides that
SG landslide detection probability. If SLDP(Oa ) is lower than were unreasonably small, while the hole filling was applied to
TSLDP , the landslide candidate object Oa would be identified fill up those small holes inside correctly detected landslides.
as a falsely detected landslide and removed. If SLDP(Oa ) is 4) Detection of Newly Occurred Landslides Based on
higher than TSLDP , the landslide candidate object Oa would be MCMO: Landslides detected by the SG-FCNN have the advan-
identified as a correctly detected landslide and retained. tages of high accuracy and reliability, making them suitable as
3) Boundary and Morphological Optimization: In the SG- the basis for detecting newly occurred landslides. As reviewed
FCNN, boundary and morphological optimization (BMO) was in Section I, the FCNN has also achieved state-of-the-art perfor-
applied to enhance the precision of detected landslides further. mance in detecting newly occurred landslides. However, the un-
In the MobileU-Net, feature maps extracted by the encoder supervised CMI-based change detection strategies have unique
had lower resolution than the input image. To obtain pixelwise advantages in efficiency [75] and are still widely applied in
landslide detection results of the same size as the input image, practical applications. Thus, combining the proposed SG-FCNN
an unpooling and convolution operation was applied to low- with the CMI-based change detection strategy is a promising
resolution feature maps in the decoder part. However, the input way to improve the detection performance of newly occurred
image was not involved in the decoder part; thus, the detected landslides.
landslides’ boundary is coarse and fails to fit the actual boundary. Based on the proposed SG-FCNN, this article further pro-
The Dense CRF [71] was applied based on aerial images’ posed an unsupervised change detection strategy, MCMO, to
spectral and textural features to refine the boundary of detected detect newly occurred landslides based on bitemporal images
landslides. Compared to the traditional CRF, which only utilizes and the landslide detection results from postlandslide images.
the connections between the target pixel and its adjacent pixels, First, the CMI is generated based on the coregistered postland-
the Dense CRF can utilize the connections between all pixel slide and prelandslide images
pairs of the input image and refine the boundary of the detected   
i abs DNi − DNi
Pre Post
landslides in greater detail. The Dense CRF confirmed to Gibbs CM (Pc ) = (8)
distribution 255 × Ni

exp (−E (X|I)) where Pc stands for the cth pixel in testing region. DNiPre and
P (x = X|I) = (4) DNiPost are the grey value of Pc in the ith band of the prelandslide
Z (I)
image and postlandslide image, respectively. The greater the
CM(Pc ), the greater the change has been experienced. Then,
where I is the input image and x is the predicted result to be
based on the changing magnitude of each pixel, the historical
optimized. E(X|I) is the Gibbs energy, which is composed of
landslides can be removed from the landslide detection results
unary potential ψu (xi ) and pairwise potential ψp (xi , yj ) as a
of postlandslide images as follows:
following formula:

  [CM (Pd )]
MCM (Oe ) = , Pd ∈ O e (9)
E (x|I) = ψu (xi ) + ψp (xi , yj ) (5) N
i i,j
where Oe stands for the eth detected landslide objects of the
ψu (xi ) = − log P (xi ) (6) postlandslide image. For Oe , CM(Pd ) is the changing magnitude
of the dth pixel, and N is the total number of pixels. Therefore,
ψp (xi , yj ) = μ (xi , yj ) MCM(Oe ) denotes the mean changing magnitude of Oe . TMCM
⎡ ⎤
is the threshold for judging whether Oe has changed or not.
⎢ 1    
|p −p |2 ⎥
|p −p |2 |Ii −Ij |2
If MCM(Oe ) exceeds TMCM , the detected landslide object Oe
×⎢ ⎣ w exp − i2θ2j − 2
2θβ
+ w2 exp − i2θ2j ⎥ ⎦ would be identified as a historical landslide or other unchanged
α γ

land cover features. TMCM was set to 25% based on trial and error.
Appearance Kernel Smoothness Kernel
Finally, the proposed MCMO strategy obtained the detection
(7)
result of newly occurred landslides by deleting those Oe with an
MCM value above TMCM .
where μ(xi , yj ) is the label compatibility. w1 and w2 are the
weights of appearance kernel potential and smoothness kernel
potential, respectively. The appearance kernel potential is used IV. EXPERIMENTS AND RESULTS
for clustering pixels with similar spectral features into the same The landslide susceptibility mapping and landslide detection
class, in which θa and θb are the degree of spatial nearness and experiments were carried out on a Windows 10 OS with a
spectral similarity. The smoothness kernel potential is applied 3.6 GHz Core i7-7700 and an NVIDIA GeForce GTX 1080.
to remove isolated regions [72] with a spatial nearness degree TensorFlow-GPU 1.7.0 was selected as the DL platform for
controlled by θγ . implementing the FCNN, the SG-FCNN, and the CPCNN-RF.
1006 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 16, 2023

Fig. 8. DAPs and evaluation samples in testing region A. (a) DAP 2007. (b) DAP 2008. (c) Evaluation samples overlaid on DAP 2008.

Fig. 9. DAPs and evaluation samples in testing region B. (a) DAP 2007. (b) DAP 2008. (c) Evaluation samples overlaid on DAP 2008.

TABLE I
STATISTICAL DETAILS OF EVALUATION SAMPLES

The Dense CRF model was implemented with Python wrapper, shown in Figs. 8 and 9. The statistical details of the evaluation
pydensecrf 1.0rc3 (https://github.com/lucasb-eyer/pydensecrf). samples are given in Table I.
There were two missions in this article: detecting all landslides
(including historical and newly occurred landslides) in 2008 and
detecting newly occurred landslides in 2008. To evaluate the A. Detection of All Landslides
detection performance in two testing regions, two ground truth During the training procedure of the FCNN, an Adam opti-
datasets were visually delineated by reference to the ENTLI mization algorithm was selected to update the networks’ weights
(from 1994 to 2008) and coregistered DAPs (2007 and 2008), as iteratively. Considering the hardware performance of the
CHEN et al.: SUSCEPTIBILITY-GUIDED LANDSLIDE DETECTION USING FULLY CONVOLUTIONAL NEURAL NETWORK 1007

experimental environment, the batch size and the initial learning

rate were set as 8 and 0.001, respectively. If the validation loss
did not decrease within three epochs, the learning rate would
reduce by 90%. The early stopping strategy was applied based on
validation loss to avoid overfitting. The network training would
be terminated if the validation loss did not decrease within 10
epochs.
As for the overlapping mosaic detection strategy, the target
area constituted 5% of the corresponding predicted patch. The
hyperparameters of the Dense CRF, namely, θa , θβ , θγ , w1 , and
w2 , were set to 30, 20, 1.0, 1.5, and 1.0 in testing region A,
and were set to 30, 10, 1.0, 1.1, and 1.0 in testing region B,
according to the suggestion of Kra¨henbu¨hl and Koltun [71]
and the optimization experience of Shi in Lantau Island [51].
The Dense CRF optimized the initial detection results for 10
iterations. Taking the characteristics of landslides and ground
dataset on Lantau Island into account, the thresholds for area Fig. 10. Landslide susceptibility map generated by CPCNN-RF based on
ENTLI (1994–2007) [31].
opening and hole filling were set as 25 and 200 m2 , respectively.
When detecting landslides using the SG-FCNN, we applied
the landslide susceptibility map generated by the state-of-the-art To highlight the superiority of the proposed SG-FCNN, we
CPCNN-RF. All details of the CPCNN-RF and susceptibility further compared it with the PBCNN, the FCNN, and the FCNN
mapping process in Lantau Island can be found in [31]. It cannot with postprocessing based on auxiliary information. The applied
present all details about the CPCNN-RF due to space limitations. PBCNN is the per superpixel multiscale CNN based on SEEDS
The CPCNN-RF was trained based on 10 landslide predispos- (SEEDS–MCNN) [26] with a MobileNet-V2 backbone. The
ing factor layers of Lantau Island and landslide crowns recorded applied FCNN includes the original U-Net (with VGG-16 back-
in the ENTLI (from 1994 to 2007). The applied landslide bone) [60], DeepLab-V3 (with MobileNet-V2 backbone) [61],
predisposing factor layers had a spatial resolution of 30 m, and the proposed MobileU-Net. The postprocessing strategy
which included elevation, aspect, slope, plan curvature, profile based on auxiliary information is referenced to Shi’s strategy for
curvature, distance to fault, lithology, topographic wetness index FCNN-based landslide detection in Lantau Island [51]: Creating
(TWI), land cover, and NDVI. Elevation, slope, aspect, profile a buffer zone based on road vectors with a distance of 10 m.
curvature, and plan curvature were calculated using the digital If the overlapping area of the landslide object and buffer zone
elevation model (DEM) (30 m per pixel) provided by Esri exceeds 30% of the landslide object, discard the corresponding
China (HK). Lithological and fault lines were extracted from the landslide object. If more than 50% area of the landslide object
simplified geological map provided by Esri China (HK). TWI with a slope less than 10°, discard the corresponding landslide
was calculated from DEM using hydrology tools in ArcGIS. object. The applied road vector of Lantau Island was downloaded
Land cover was obtained from the FROM-GLC product (30 m from https://www.openstreetmap.org. The slope map of Lantau
per pixel) [76]. NDVI was calculated using Landsat TM images Island was calculated based on the DEM with a spatial resolution
with 30-m spatial resolution, which was taken on 14th March of 5 m, provided by ESRI Hong Kong. All comparative mod-
2006. After training ended, the predisposing factor layers of els mentioned above were pretrained by the ImageNet image
Lantau Island were divided into 163 052 grids with a spatial classification dataset and applied the training data augmentation
resolution of 30 m and fed into the trained CPCNN-RF for strategy mentioned in Section III-A-2. All FCNN-based compar-
landslide susceptibility mapping. The derived landslide suscep- ative models applied the overlapping mosaic detection strategy
tibility map is given in Fig. 10, which has been ranked into five mentioned in Section III-B-1.
grades.
Landslide crowns recorded in the ENTLI (from 1994 to 2007)
were used as testing samples to evaluate the reliability of the B. Detection of Newly Occurred Landslides
derived landslide susceptibility map. The OA and the area under Landslides detected by the SG-FCNN in the DAPs (2008)
the receiver operating characteristic curve were determined in were used as the basis for detecting newly occurred landslides
SPSS Statistic 26 as statistical evaluation metrics. The evaluation in 2008. The CMI of the two testing areas was generated based on
results illustrated that the landslide susceptibility map generated the coregistered DAPs (2007) and DAPs (2008), using ArcMap
by the CPCNN-RF has achieved satisfactory accuracy, which 10.3. Figs. 15 and 16 illustrate the intermediate process of the
qualified it to be used as an indicator to distinguish the areas SG-FCNN with MCMO strategy in testing region A and B.
that were not prone to landslide occurrence. Figs. 17 and 18 represent the detection results of newly occurred
For the two testing regions, Figs. 11 and 12 show the inter- landslides in 2008.
mediate process of the SG-FCNN. Figs. 13 and 14 show the The proposed MCMO change detection strategy was
detection results of all landslides in 2008. compared with the change detection strategy applied by the
1008 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 16, 2023

Fig. 11. Intermediate process of SG-FCNN in testing region A. (a) Landslide susceptibility. (b) Landslide detection probability. (c) Susceptibility-guided landslide
detection probability.

Fig. 12. Intermediate process of SG-FCNN in testing region B. (a) Landslide susceptibility. (b) Landslide detection probability. (c) Susceptibility-guided landslide
detection probability.

object-oriented change detection CNN (CDCNN) [51]. In the C. Evaluation of Detection Results
CDCNN, FCNN is used to detect landslides from prelandslide Four widely used metrics, including precision, recall, F1
and postlandslide images separately. Based on this, newly score, and IoU were applied and can be defined by the following
occurred landslides are identified by analyzing the IoU of
equations:
paired landslide objects detected from the prelandslide and
postlandslide images. If a landslide object detected in the
postlandslide image failed to pair with a landslide object PNa
Precision = (10)
in the prelandslide image, or the IoU of paired landslide PNb
objects is less than 50%, the corresponding landslide is newly PNa
occurred. Recall = (11)
PNc
CHEN et al.: SUSCEPTIBILITY-GUIDED LANDSLIDE DETECTION USING FULLY CONVOLUTIONAL NEURAL NETWORK 1009

Fig. 13. Detection results of all landslides in testing region A. (a) Ground truth. (b) PBCNN (SEEDS–MCNN) [26]. (c) FCNN (original U-Net) [60]. (d) FCNN
(DeepLab-V3) [61]. (e) FCNN (MobileU-Net). (f) FCNN (MobileU-Net) + PA-TTA. (g) FCNN (MobileU-Net) + BMO. (h) FCNN (MobileU-Net) + BMO +
post-processing strategy based on auxiliary information [51]. (i) SG-FCNN (MobileU-Net) + BMO.
1010 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 16, 2023

Fig. 14. Detection results of all landslides in testing region B. (a) Ground truth. (b) PBCNN (SEEDS-MCNN) [26]. (c) FCNN (original U-Net) [60]. (d) FCNN
(DeepLab-V3) [61]. (e) FCNN (MobileU-Net). (f) FCNN (MobileU-Net) + PA-TTA. (g) FCNN (MobileU-Net) + BMO. (h) FCNN (MobileU-Net) + BMO +
postprocessing strategy based on auxiliary information [51]. (i) SG-FCNN (MobileU-Net) + BMO.

2 × Precision × Recall of correctly predicted landslide pixels in predicted landslide

F1 Score = (12)
Precision + Recall pixels. Recall measures the proportion of landslide pixels that
are correctly detected. F1 Score and IoU are both comprehensive
PNa indexes. The F1 Score is the harmonic mean of precision and
IoU = (13)
PNc + PNd recall, and the IoU measures the extent of the overlap of detected
landslides and ground truth. The evaluation results of the two
where PNa stands for the pixel number of correctly detected testing regions are given in Tables II and III.
landslides. PNb is the pixel number of detected landslides. PNc While detecting all landslides in 2008, the FCNN was superior
is stands for the pixel number of landslide ground truth. PNd is to the PBCNN. The detection accuracy of the MobileU-Net was
the pixel number of detected landslide that is not matched with much higher than that of the original U-Net and is comparable
corresponding ground truth. Precision measures the proportion to that of the MobileU-Net. The MobileU-Net only have 6
CHEN et al.: SUSCEPTIBILITY-GUIDED LANDSLIDE DETECTION USING FULLY CONVOLUTIONAL NEURAL NETWORK 1011

Fig. 15. Intermediate process of the SG-FCNN with MCMO in testing region A. (a) Changing magnitude. (b) Mean changing magnitude.

Fig. 16. Intermediate process of the SG-FCNN with MCMO in testing region B. (a) Changing magnitude. (b) Mean changing magnitude.

315 522 parameters, which is only half of that of the original While detecting newly occurred landslides in 2008, the de-
U-Net and one-third of that of the DeepLab-V3. Thus, the tection accuracy of the SG-FCNN with the proposed MCMO
MobileU-Net is much more efficient and has the best compre- change detection strategy was significantly improved compared
hensive performance among all FCNN models. With the help with the SG-FCNN with the change detection strategy of CD-
of the susceptibility guidance strategy, the detection accuracy of CNN [51]. In addition, no matter what change detection strategy
the proposed SG-FCNN was improved significantly compared was combined with the proposed SG-FCNN, the detection accu-
to the original FCNN. In particular, the SG-FCNN with BMO racy of newly occurred landslides was higher than the detection
achieved the outstanding performance with the highest pre- accuracy of all landslides. Especially in testing region B, by
cision, F1 Score, and IoU in all testing regions. However, a utilizing the proposed MCMO change detection strategy, the
small number of land cover features were falsely identified as detection accuracy of newly occurred landslides was sharply
landslides by the SG-FCNN. improved compared with the detection accuracy of all landslides.
1012 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 16, 2023

Fig. 17. Detection results of newly occurred landslides in testing region A. (a) Ground truth of newly occurred landslides in 2008. (b) SG-FCNN + change
detection strategy of CDCNN [51]. (c) SG-FCNN + MCMO change detection strategy.

Fig. 18. Detection results of newly occurred landslides in testing region B. (a) Ground truth of newly occurred landslides in 2008. (b) SG-FCNN + change
detection strategy of CDCNN [51]. (c) SG-FCNN + MCMO change detection strategy.

V. DISCUSSION FP-A2 in Fig. 13(g), FP-B1 to FP-B8 in Fig. 14(g)]. Tables II and
III show that the SG-FCNN is superior to the original MobileU-
A. Effectiveness of Susceptibility Guidance Strategy
Net and the MobileU-Net with PA-TTA on all evaluation indices.
This article modified the FCNN and innovatively proposed the In testing region A, the F1 Score and IoU of the SG-FCNN
SG-FCNN by introducing the susceptibility guidance strategy. reached 0.9049 and 82.63%, respectively. In testing region B,
For each landslide candidate object, its averaged probability the SG-FCNN achieved an F1 Score of 0.7938 and an IoU of
of landslide detection merged with the landslide susceptible 65.81%.
probability. Compared to the state-of-the-art postprocessing method based
In the proposed SG-FCNN, visual and environmental condi- on auxiliary information (road vectors, building vectors, and
tions would be considered simultaneously during the detection slope) [51], the proposed SG-FCNN also has a clear lead on
process. Landslide candidates with low visual detection prob- all evaluation indices. Compared to the SG-FCNN, FP-A3 in
ability in high landslide susceptibility zones can be detected Fig. 13(h) and FP-B9 to FP-B14 in Fig. 14(h) were falsely
[see TP-A5 and TP-A6 in Fig. 13(i)]. In contrast, landslide detected, TP-A5 to TP-A10 in Fig. 13(i) and TP-B5 to TP-B7 in
candidates with high detection visual probability in low Fig. 14(i) were not detected. The reason for the success of the
landslide susceptibility zones can be discarded [see FP-A1 and proposed SG-FCNN is that it extracted knowledge (landslide
CHEN et al.: SUSCEPTIBILITY-GUIDED LANDSLIDE DETECTION USING FULLY CONVOLUTIONAL NEURAL NETWORK 1013

TABLE II
QUANTITATIVE EVALUATION RESULTS OF LANDSLIDE DETECTION RESULTS IN TESTING REGION A

TABLE III
QUANTITATIVE EVALUATION RESULTS OF LANDSLIDE DETECTION RESULTS IN TESTING REGION B

susceptibility) from auxiliary information (landslide predispos- B. Effectiveness and Indispensability of PA-TTA and
ing factors and the distribution of historical landsides) instead Susceptibility Guidance
of using auxiliary information directly. As prior knowledge,
landslide susceptibility condenses assorted auxiliary informa- While aiming to improve the detection accuracy of the FCNN,
tion related to landslide occurrence and considers the occurrence this article proposed an SG-FCNN that used PA-TTA and inno-
rule of historical landslides simultaneously. Given the above, vatively introduced landslide susceptibility as prior knowledge
introducing landslide susceptibility can significantly improve for landslide detection. After the FCNN initially detected the
the detection accuracy of FCNN and expand the application landslides, the PA-TTA was applied, and landslide susceptibility
range of landslide susceptibility. was involved in guiding the detection process. Misdetections
1014 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 16, 2023

were avoided, and false detections were eliminated, thus, adding by the SG-FCNN in testing region A is already satisfactory,
to the accuracy of the landslide detection. and the number of falsely detected landslides is scarce. This
There are three reasons for applying the PA-TTA rather than also explains why the accuracy increase in testing region A
the max voting test time augmentation (MV-TTA). First, the is lower than in testing region B after applying the proposed
MV-TTA has few advantages in landslide detection. As shown MCMO strategy.
in Tables II and III, the detection accuracy of the MobileU-Net In this article, the proposed MCMO was compared with
slightly increased after applying the MV-TTA. Second, the PA- the state-of-the-art change detection strategy of the CDCNN
TTA is implemented based on the probability map instead of the developed by Shi et al. [51]. As presented in Tables II, III,
binary detection result. Thus, a generated average probability Figs. 17(b), and 18(b), the change detection strategy of the
map of landslide detection can be analyzed with a landslide CDCNN has achieved a certain effect. Almost all historical
susceptibility map, and SG landslide detection can be realized. landslides were identified and removed from the detection results
Third, the proposed PA-TTA only discarded those pixels with of the SG-FCNN. In comparison, the proposed MCMO strategy
very low averaged probability and can detect more landslide is more effective with much higher accuracy in the two testing
candidates than the original FCNN with or without the MV-TTA. regions. The MCMO performs better as all unchanged land
As presented in Table II and III, applying the PA-TTA can effec- cover features, including historical landslides, can be identified
tively avoid the misdetection of landslides with a significantly and removed. In contrast, the change detection strategy of the
higher recall. CDCNN only identified and removed historical landslides but
Compared with applying the MV-TTA, applying the PA-TTA ignored other unchanged land cover features [FP-A12 to FP-A16
would result in lower precision and high recall rate because in Fig. 17(b), FP-B22 to FP-B26 in Fig. 18(b)]. In testing
the probability threshold of the PA-TTA is set as 20% in- region B, some newly occurred landslides were misidentified
stead of 50% to detect landslide candidates as much as pos- as historical landslides and removed by the change detection
sible. But most falsely detected landslides can be removed strategy of the CDCNN [see FP-B27 and FP-B28 in Fig. 18(b)].
with ensued joint analysis guided by landslide susceptibility. In addition, applying the MCMO strategy only need to calculate
Thus, it can be concluded that the PA-TTA and the suscep- the mean changing magnitude of detected landslides based on
tibility guidance strategy are interdependent and indispens- CMI with a simple operation, and the whole processing time is
able. Without PA-TTA, applying the susceptibility guidance less than a second. In contrast, the change detection strategy of
strategy would result in a low recall rate. Without the sus- the CDCNN needs to carry out an additional landslide detection
ceptibility guidance, applying the PA-TTA would not achieve process of prelandslide images using the SG-FCNN or FCNN,
high precision. which significantly increases the processing time and computa-
tion burden. Thus, the MCMO strategy is much more efficient
and easier to implement.
C. Effectiveness of Combining SG-FCNN With MCMO Given the above, with the help of the proposed MCMO,
As discussed above, the original SG-FCNN can detect all the proposed SG-FCNN has better accuracy and efficiency in
landslides (both historical and newly occurred landslides) from detecting newly occurred landslides.
single temporal DAPs with high precision. To further expand
the application scope of the SG-FCNN, an unsupervised change
detection strategy MCMO was proposed to detect newly oc- D. Defects of Proposed Landslide Detection Method
curred landslides from bitemporal DAPs. As shown in Tables II Despite the advantages of the proposed SG-FCNN and
and III, after integrating with the proposed MCMO strategy, the MCMO, false and misdetection of landslides cannot be avoided
SG-FCNN achieved significantly higher accuracy in detecting completely. A small number of muddy tracks, stony creeks,
newly occurred landslides than the original SG-FCNN. The and other surface exposures not caused by landslides [FP-A4
result is especially obvious in testing region B, where the F1 to FP-A11 in Fig. 13(i), FP-B15 to FP-B21 in Fig. 14(i)] were
Score and IoU increased by 0.07 and 10.22%, reaching 0.8638 falsely detected using the SG-FCNN due to the lack of assorted
and 76.03%, respectively. The detection accuracy of testing negative samples. In addition, the proposed method failed to
region A is much higher, with an F1 Score of 0.9169 and detect those small landslides with a large length–width ratio
IoU of 84.65%, but the accuracy only increased by 0.012 and because it is difficult to distinguish them from narrow roads
2.02%, respectively. As shown in Figs. 17(c) and 18(c), the based only on visual features [TP-A1 to TP-A4 in Fig. 13(a),
SG-FCNN successfully identified newly occurred landslides TP-B1 to TP-B4 in Fig. 14(a)]. Although the proposed MCMO
from all landslides with the help of the proposed MCMO. Espe- strategy can identify most of the falsely detected landslides
cially in testing region B, numerous falsely detected landslides as unchanged land cover features and then discard them, the
from a single temporal image, such as stony creeks [FP-B15 others were still falsely detected because of the coregistration
to FP-B18 in Fig. 14(i)], bare rocks [FP-B21 in Fig. 14(i)] and error of bitemporal images away from the isocenter [FP-A17
muddy tracks [FP-B21 in Fig. 14(i)] has also been identified in Fig. 17(c)] and surface changes not caused by landslides
as unchanged objects and removed. Compared to testing region [FP-A18 in Fig. 17(c), FP-B29 and FP-B30 in Fig. 18(c)].
B, the land-cover composition in testing region A is relatively Meanwhile, a small amount of newly occurred landslides were
simple, and the visual characteristics of landslides are clear falsely identified as unchanged objects and removed [TP-B8
and distinguishable. Thus, the detection result of all landslides and TP-B9 in Fig. 18(a)] because their spectral characteristics
CHEN et al.: SUSCEPTIBILITY-GUIDED LANDSLIDE DETECTION USING FULLY CONVOLUTIONAL NEURAL NETWORK 1015

were not changed significantly. Introducing nonoptical remote but also presents a new methodology for landslide detection
sensing products and high-level feature extractions methods on by introducing landslide susceptibility as knowledge guidance,
a multiscale should be considered in future research. which is considered more important. The application range of
Besides, the proposed SG-FCNN would result in additional landslide susceptibility maps would not be limited to landslide
time cost and parameter amount compared with the FCNN. The risk management and can be adopted in landslide inventory
PA-TTA strategy requires the FCNN to detect the landslides mapping. Considering that this is the first attempt to integrate
multiple times. Thus, the additional time cost caused is obvious. landslide susceptibility and detection, this article also has a
The specific time cost depends on the number of applied data certain insufficiency as discussed in Section V-D, which should
augmentation methods. In contrast, the PA-TTA strategy will not be addressed in our future article.
introduce new parameters. As for the SG strategy, the additional
cost and parameter amount depend on the complexity of the
ACKNOWLEDGMENT
applied landslide susceptibility mapping model. We will try to
further improve the efficiency of the SG-FCNN by utilizing more We would like to thank Mr. Wai Kin Leung, the Geotechnical
efficient data-augmentation methods and landslide susceptibility Engineer from CEDD of the Hong Kong Special Administrative
model in our future article. Region, for providing the ENTLI dataset. We appreciate Dr.
In addition, although the applied CPCNN-RF has achieved Min Zhang from the Hong Kong Polytechnic University for his
the best performance in Lantau Island, the generated landslide valuable suggestions on the preparatory work of experiments.
susceptibility map still could not guarantee 100% accuracy of We also thank the editors and anonymous reviewers for their
landslide occurrence. The errors of the landslide susceptibility constructive comments.
mapping would be transferred to landslide detection in the
proposed SG-FCNN. Thus, it is equally important to study how
REFERENCES
to improve the reliability of landslide susceptibility mapping and
reduce the negative impact caused by error transmission in our [1] X. Deng, D. Xu, M. Zeng, and Y. Qi, “Landslides and cropland aban-
donment in China’s mountainous areas: Spatial distribution, empirical
future article. Furthermore, limited by the quality of the data analysis and policy implications,” Sustainability, vol. 10, no. 11, 2018,
source, the spatial resolution of the landslide susceptibility map Art. no. 3909.
generated by the CPCNN-RF is only 30 m, which is inconsistent [2] C. L. H. Lam, J. W. C. Lau, and H. W. Chan, “Factual report on Hong
Kong rainfall and landslides in 2008,” Geotech. Eng. Office, Hong Kong,
with the spatial resolution of DAPs. Applying a fine-grained GEO Report No. 273, 2009.
landslide susceptibility map is a promising direction for min- [3] B.-G. Chae, H.-J. Park, F. Catani, A. Simoni, and M. Berti, “Landslide
imizing the error caused by data scaling and improving the prediction, monitoring and early warning: A concise review of state-of-
the-art,” Geosci. J., vol. 21, no. 6, pp. 1033–1070, 2017.
detection performance further. [4] M. Galli, F. Ardizzone, M. Cardinali, F. Guzzetti, and P. Reichenbach,
“Comparing landslide inventory maps,” Geomorphology, vol. 94, no. 3/4,
pp. 268–289, 2008.
VI. CONCLUSION [5] C. J. Van Westen, E. Castellanos, and S. L. Kuriakose, “Spatial data
for landslide susceptibility, hazard, and vulnerability assessment: An
This article took a high-incidence area of landslide in Lantau overview,” Eng. Geol., vol. 102, no. 3/4, pp. 112–131, 2008.
Island as the study area and researched FCNN-based landslide [6] F. Guzzetti, A. C. Mondini, M. Cardinali, F. Fiorucci, M. Santangelo, and
K.-T. Chang, “Landslide inventory maps: New tools for an old problem,”
detection method. First, various algorithm optimization methods Earth-Sci. Rev., vol. 112, no. 1-2, pp. 42–66, 2012.
such as overlapping mosaic detection strategy, compound loss [7] C. Van Westen, T. W. Van Asch, and R. Soeters, “Landslide hazard and risk
function, and dense CRF were applied, which improved the zonation—Why is it still so difficult?,” Bull. Eng. Geol. Environ., vol. 65,
no. 2, pp. 167–184, 2006.
detection accuracy of FCNN toward landslides’ boundary. Based [8] D. Brunsden, “Landslide types, mechanisms, recognition, identification,”
on this, the SG-FCNN was proposed to detect landslides under Landslides in the South Wales Coalfield, Edited by: Morgan, CS, 1985,
the guidance of landslide susceptibility. Through the analysis pp. 1–3.
[9] A. Xing et al., “Dynamic analysis and field investigation of a fluidized
of experimental results, it can be concluded that the proposed landslide in Guanling, Guizhou, China,” Eng. Geol., vol. 181, pp. 1–14,
landslide susceptibility guidance strategy has shown great po- 2014.
tential in detecting landslides. By introducing landslide suscepti- [10] G. Metternicht, L. Hurni, and R. Gogu, “Remote sensing of landslides:
An analysis of the potential contribution to geo-spatial systems for hazard
bility as prior knowledge, the proposed SG-FCNN significantly assessment in mountainous environments,” Remote Sens. Environ., vol. 98,
outperforms the FCNN and the state-of-the-art post-processing no. 2/3, pp. 284–303, 2005.
method based on auxiliary information, with fewer false de- [11] J. E. Nichol, A. Shaker, and M.-S. Wong, “Application of high-resolution
stereo satellite images to detailed landslide hazard assessment,” Geomor-
tection. Besides, an unsupervised change detection strategy, phology, vol. 76, no. 1/2, pp. 68–75, 2006.
MCMO, was proposed and integrated with the SG-FCNN to [12] X. X. Zhu et al., “Deep learning in remote sensing: A comprehensive
detect newly occurred landslides based on bitemporal VHR review and list of resources,” IEEE Geosci. Remote Sens. Mag., vol. 5,
no. 4, pp. 8–36, Dec. 2017.
remote sensing images. The proposed MCMO strategy is easy to [13] F. Catani, “Landslide detection by deep learning of non-nadiral and crowd-
implement and superior to the state-of-the-art change detection sourced optical images,” Landslides, vol. 18, no. 3, pp. 1025–1044, 2021.
technique in terms of detection accuracy and efficiency. Apply- [14] Z. Li, W. Shi, P. Lu, L. Yan, Q. Wang, and Z. Miao, “Landslide
mapping from aerial photographs using change detection-based Markov
ing the MCMO can fully utilize the merits of the unsupervised random field,” Remote Sens. Environ., vol. 187, pp. 76–90, 2016,
change detection strategy and the proposed SG-FCNN. doi: 10.1016/j.rse.2016.10.008.
Given the above, the proposed SG-FCNN and MCMO strat- [15] J. G. Speight, “Landform pattern description from aerial photographs,”
Photogrammetria, vol. 32, no. 5, pp. 161–182, 1977.
egy raises the performance of landslide detection to a new level
1016 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 16, 2023

[16] H. Cai, T. Chen, R. Niu, and A. Plaza, “Landslide detection using densely [38] M. I. Sameen and B. Pradhan, “Landslide detection using residual net-
connected convolutional networks and environmental conditions,” IEEE works and the fusion of spectral and topographic information,” IEEE
J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, pp. 5235–5247, Access, vol. 7, pp. 114363–114373, 2019.
May 2021. [39] S. Ji, D. Yu, C. Shen, W. Li, and Q. Xu, “Landslide detection from an
[17] G. Danneels, E. Pirard, and H.-B. Havenith, “Automatic landslide detection open satellite imagery and digital elevation model dataset using atten-
from remote sensing images using supervised classification methods,” in tion boosted convolutional neural networks,” Landslides, vol. 17, no. 6,
Proc. IEEE Int. Geosci. Remote Sens. Symp., 2007, pp. 3014–3017. pp. 1337–1352, 2020.
[18] V. Moosavi, A. Talebi, and B. Shirmohammadi, “Producing a landslide in- [40] L. Bragagnolo, L. R. Rezende, R. V. da Silva, and J. M. V. Grzy-
ventory map using pixel-based and object-oriented approaches optimized bowski, “Convolutional neural networks applied to semantic segmen-
by Taguchi method,” Geomorphology, vol. 204, pp. 646–656, 2014. tation of landslide scars,” Catena, vol. 201, 2021, Art. no. 105189,
[19] D. T. Bui et al., “Landslide detection and susceptibility mapping by airsar doi: 10.1016/j.catena.2021.105189.
data using support vector machine and index of entropy models in Cameron [41] S. R. Meena et al., “Landslide detection in the Himalayas using machine
highlands, Malaysia,” Remote Sens., vol. 10, no. 10, 2018, Art. no. 1527. learning algorithms and U-Net,” Landslides, vol. 19, no. 5, pp. 1209–1229,
[20] A. Stumpf and N. Kerle, “Object-oriented mapping of landslides using 2022.
random forests,” Remote Sens. Environ., vol. 115, no. 10, pp. 2564–2577, [42] N. Prakash, A. Manconi, and S. Loew, “Mapping landslides on EO data:
2011. Performance of deep learning models vs. traditional machine learning
[21] W. Zhao and S. Du, “Learning multiscale and deep representations for models,” Remote Sens., vol. 12, no. 3, 2020, Art. no. 346.
classifying remotely sensed imagery,” ISPRS J. Photogramm. Remote [43] P. Liu, Y. Wei, Q. Wang, Y. Chen, and J. Xie, “Research on post-earthquake
Sens., vol. 113, pp. 155–165, 2016. landslide extraction algorithm based on improved u-net model,” Remote
[22] C. Zhong et al., “Landslide mapping with remote sensing: Challenges and Sens., vol. 12, no. 5, 2020, Art. no. 894.
opportunities,” Int. J. Remote Sens., vol. 41, no. 4, pp. 1555–1581, 2020. [44] O. Ghorbanzadeh, K. Gholamnia, and P. Ghamisi, “The application of
[23] L. Wan, T. Chen, A. Plaza, and H. Cai, “Hyperspectral unmixing based ResU-Net and OBIA for landslide detection from multi-temporal sentinel-
on spectral and sparse deep convolutional neural networks,” IEEE J. 2 images,” Big Earth Data, pp. 1–26, 2022.
Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, pp. 11669–11682, [45] P. Liu et al., “A research on landslides automatic extraction model based
Nov. 2021. on the improved mask R-CNN,” ISPRS Int. J. Geo-Inf., vol. 10, no. 3,
[24] H. Larochelle, Y. Bengio, J. Louradour, and P. Lamblin, “Exploring 2021, Art. no. 168.
strategies for training deep neural networks,” J. Mach. Learn. Res., vol. 10, [46] O. Ghorbanzadeh, S. R. Meena, H. S. S. Abadi, S. T. Piralilou, L. Zhiy-
pp. 1–40, 2009. ong, and T. Blaschke, “Landslide mapping using two main deep-learning
[25] W. Zhou, Z. Shao, and Q. Cheng, “Deep feature representations for high- convolution neural network streams combined by the Dempster–Shafer
resolution remote sensing scene classification,” in Proc. 4th Int. Workshop model,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14,
Earth Observ. Remote Sens. Appl., 2016, pp. 338–342. pp. 452–463, Jan. 2021.
[26] Y. Chen, D. Ming, and X. Lv, “Superpixel based land cover classification [47] T. Lei, Y. Zhang, Z. Lv, S. Li, S. Liu, and A. K. Nandi, “Landslide in-
of VHR satellite image combining multi-scale CNN and scale parameter ventory mapping from bitemporal images using deep convolutional neural
estimation,” Earth Sci. Inform., vol. 12, no. 3, pp. 341–363, 2019. networks,” IEEE Geosci. Remote Sens. Lett., vol. 16, no. 6, pp. 982–986,
[27] D. He, Q. Shi, X. Liu, Y. Zhong, and X. Zhang, “Deep subpixel map- Jun. 2019.
ping based on semantic information modulated network for urban land [48] X. Gao, T. Chen, R. Niu, and A. Plaza, “Recognition and mapping of land-
use mapping,” IEEE Trans. Geosci. Remote Sens., vol. 59, no. 12, slide using a fully convolutional DenseNet and influencing factors,” IEEE
pp. 10628–10646, Dec. 2021. J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, pp. 7881–7894,
[28] X. Li, M. He, H. Li, and H. Shen, “A combined loss-based multiscale Aug. 2021.
fully convolutional network for high-resolution remote sensing image [49] T. Liu, T. Chen, R. Niu, and A. Plaza, “Landslide detection mapping
change detection,” IEEE Geosci. Remote Sens. Lett., vol. 19, Jan. 2022, employing CNN, ResNet, and DenseNet in the Three Gorges Reservoir,
Art. no. 8017505. China,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14,
[29] Q. Shi, M. Liu, S. Li, X. Liu, F. Wang, and L. Zhang, “A deeply supervised pp. 11417–11428, Nov. 2021.
attention metric-based network and an open aerial image dataset for remote [50] O. Ghorbanzadeh, H. Shahabi, A. Crivellari, S. Homayouni, T. Blaschke,
sensing change detection,” IEEE Trans. Geosci. Remote Sens., vol. 60, and P. Ghamisi, “Landslide detection using deep learning and object-based
2021, Art. no. 5604816. image analysis,” Landslides, vol. 19, no. 4, pp. 929–939, 2022.
[30] W. Wu et al., “Application of local fully convolutional neural network [51] W. Shi, M. Zhang, H. Ke, X. Fang, Z. Zhan, and S. Chen, “Landslide recog-
combined with YOLO v5 algorithm in small target detection of remote nition by deep convolutional neural network and change detection,” IEEE
sensing image,” PLoS One, vol. 16, no. 10, 2021, Art. no. e0259283. Trans. Geosci. Remote Sens., vol. 59, no. 6, pp. 4654–4672, Jun. 2021.
[31] Y. Chen, D. Ming, X. Ling, X. Lv, and C. Zhou, “Landslide susceptibility [52] K. S. Cheng, C. Wei, and S. Chang, “Locating landslides using multi-
mapping using feature fusion-based CPCNN-ML in Lantau Island, Hong temporal satellite images,” Adv. Space Res., vol. 33, no. 3, pp. 296–301,
Kong,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, 2004.
pp. 3625–3639, Apr. 2021. [53] D. Wen et al., “Change detection from very-high-spatial-resolution optical
[32] L. Lv, T. Chen, J. Dou, and A. Plaza, “A hybrid ensemble-based deep- remote sensing images: Methods, applications, and future directions,”
learning framework for landslide susceptibility mapping,” Int. J. Appl. IEEE Geosci. Remote Sens. Mag., vol. 9, no. 4, pp. 68–101, Dec. 2021,
Earth Observ. Geo-Inf., vol. 108, 2022, Art. no. 102713. doi: 10.1109/mgrs.2021.3063465.
[33] Z. Ma, G. Mei, E. Prezioso, Z. Zhang, and N. Xu, “A deep learning [54] P. Reichenbach, M. Rossi, B. D. Malamud, M. Mihir, and F. Guzzetti, “A
approach using graph convolutional networks for slope deformation pre- review of statistically-based landslide susceptibility models,” Earth-Sci.
diction based on time-series displacement data,” Neural Comput. Appl., Rev., vol. 180, pp. 60–91, 2018.
vol. 33, no. 21, pp. 14441–14457, 2021. [55] F. Guzzetti, “Landslide hazard and risk assessment,” Ph.D. dissertation,
[34] Q. Meng, H. Wang, M. He, J. Gu, J. Qi, and L. Yang, “Displacement Univ. Bonn (Rheinische Friedrich-Wilhelms-Universität Bonn), Bonn,
prediction of water-induced landslides using a recurrent deep learning Germany, 2006.
model,” Eur. J. Environ. Civil Eng., pp. 1–15, 2020. [56] R. Langford, J. James, R. Shaw, S. Campbell, P. Kirk, and R. Sewell,
[35] O. Ghorbanzadeh, A. Crivellari, P. Ghamisi, H. Shahabi, and T. Blaschke, Geology of Lantau District, vol. 6. Hong Kong: Geotechnical Eng. Office
“A comprehensive transferability evaluation of U-Net and ResU-Net for Hong Kong Government, 1995.
landslide detection from sentinel-2 data (case study areas from Taiwan, [57] “Climate change and extreme landslide events,” G. E. Office, 2021.
China, and Japan),” Sci. Rep., vol. 11, no. 1, Jul. 2021, Art. no. 14629, [58] Z. Su, J. K. Chow, P. S. Tan, J. Wu, Y. K. Ho, and Y.-H. Wang, “Deep con-
doi: 10.1038/s41598-021-94190-9. volutional neural network–based pixel-wise landslide inventory mapping,”
[36] H. Yu, Y. Ma, L. Wang, Y. Zhai, and X. Wang, “A landslide intelligent Landslides, vol. 18, no. 4, pp. 1421–1443, 2021.
detection method based on CNN and RSG_R,” in Proc. IEEE Int. Conf. [59] M.-F. J. Venture, “Final report on compilation of the enhanced natural
Mechatron. Automat., 2017, pp. 40–44. terrain landslide inventory (ENTLI),” Maunsell-Fugro Joint Venture and
[37] O. Ghorbanzadeh, T. Blaschke, K. Gholamnia, S. R. Meena, D. Tiede, and Geotechnical Engineering Office, Hong Kong, 2007.
J. Aryal, “Evaluation of different machine learning methods and deep- [60] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks
learning convolutional neural networks for landslide detection,” Remote for biomedical image segmentation,” in Proc. Int. Conf. Med. Image
Sens., vol. 11, no. 2, 2019, Art. no. 196. Comput. Comput.-Assisted Intervention, 2015, pp. 234–241.
CHEN et al.: SUSCEPTIBILITY-GUIDED LANDSLIDE DETECTION USING FULLY CONVOLUTIONAL NEURAL NETWORK 1017

[61] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mo- Dongping Ming (Member, IEEE) received the B.E.
bilenetv2: Inverted residuals and linear bottlenecks,” in Proc. IEEE Conf. degree in land administration and cadastral surveying
Comput. Vis. Pattern Recognit., 2018, pp. 4510–4520. from the Wuhan Technical University of Surveying
[62] A. Odena, V. Dumoulin, and C. Olah, “Deconvolution and checkerboard and Mapping, Wuhan, China, in 1999, the M.E. de-
artifacts,” 2016. [Online]. Available: https://distill.pub/2016/deconv- gree in cartography and geographic information en-
checkerboard/ gineering from the Wuhan University, Wuhan, China,
[63] F. Milletari, N. Navab, and S.-A. Ahmadi, “V-net: Fully convolutional in 2002, and the Ph.D. degree in cartography and geo-
neural networks for volumetric medical image segmentation,” in Proc. 4th graphic information system from the Institute of Geo-
Int. Conf. 3D Vis., 2016, pp. 565–571. graphical Sciences and Natural Resources Research,
[64] Y. Ho and S. Wookey, “The real-world-weight cross-entropy loss function: Chinese Academy of Sciences, Beijing, China, in
Modeling the costs of mislabeling,” IEEE Access, vol. 8, pp. 4806–4813, 2006.
2020. She is currently a Professor with the School of Information Engineering,
[65] E. W. Weisstein, “Moore neighborhood,” From MathWorld–A Wolfram China University of Geosciences, Beijing. Her research interests include remote-
Web Resource, 2005. [Online]. Available: http://mathworld.wolfram.com/ sensing image processing and analysis, information extraction from high spa-
MooreNeighborhood.html tial resolution satellite remote sensing images, and spatial statistics and its
[66] E. E. Brabb, “Innovative approaches to landslide hazard and risk mapping,” application.
in Proc. Int. Landslide Symp., 1985, pp. 17–22.
[67] S. D. Pardeshi, S. E. Autade, and S. S. Pardeshi, “Landslide hazard
assessment: Recent trends and techniques,” Springer Plus, vol. 2, no. 1,
pp. 1–11, 2013.
[68] Y. Wang, Z. Fang, and H. Hong, “Comparison of convolutional neural
networks for landslide susceptibility mapping in Yanshan County, China,”
Sci. Total Environ., vol. 666, pp. 975–993, 2019.
[69] Z. Fang, Y. Wang, L. Peng, and H. Hong, “Integration of convolutional
neural network and conventional machine learning classifiers for landslide
susceptibility mapping,” Comput. Geosci., vol. 139, 2020, Art. no. 104470.
[70] H. Wang, L. Zhang, H. Luo, J. He, and R. Cheung, “AI-powered landslide Junchuan Yu received the Ph.D. degree in miner-
susceptibility assessment in Hong Kong,” Eng. Geol., vol. 288, 2021, alogy, petrology, mineral deposit geology from the
Art. no. 106103. China University of Geosciences, Beijing, China, in
[71] P. Krähenbühl and V. Koltun, “Efficient inference in fully connected CRFs 2013.
with Gaussian edge potentials,” Adv. Neural Inf. Process. Syst., vol. 24, He is currently working with the Department of
pp. 109–117, 2011. Satellite Application Research, China Aero Geophys-
[72] J. Shotton, J. Winn, C. Rother, and A. Criminisi, “Textonboost for image ical Survey and Remote Sensing Center for Natural
understanding: Multi-class object recognition and segmentation by jointly Resources, Beijing, China. His research interests in-
modeling texture, layout, and context,” Int. J. Comput. Vis., vol. 81, no. 1, clude hyperspectral remote sensing, deep learning,
pp. 2–23, 2009. and geological application.
[73] P. Soille, Morphological Image Analysis: Principles and Applications.
Berlin, Germany: Springer-Verlag, 1999.
[74] Z. Li, W. Shi, S. W. Myint, P. Lu, and Q. Wang, “Semi-automated
landslide inventory mapping from bitemporal aerial photographs using
change detection and level set method,” Remote Sens. Environ., vol. 175,
pp. 215–230, 2016.
[75] Z. Y. Lv, W. Shi, X. Zhang, and J. A. Benediktsson, “Landslide inven-
tory mapping from bitemporal high-resolution remote sensing images
using change detection and multiscale segmentation,” IEEE J. Sel. Top-
ics Appl. Earth Observ. Remote Sens., vol. 11, no. 5, pp. 1520–1532, Lu Xu received the B.S. degree in geographical
May 2018. information science from the China University of
[76] P. Gong et al., “Finer resolution observation and monitoring of global land Geosciences, Beijing, China, in 2018. He is currently
cover: First mapping results with Landsat TM and ETM+ data,” Int. J. working toward the Ph.D. degree in surveying and
Remote Sens., vol. 34, no. 7, pp. 2607–2654, 2013. mapping with the China University of Geosciences.
His research interests include theory of object-
based image analysis and image information extrac-
tion by deep-learning methods.

Yangyang Chen was born in Beijing, China, in

1992. He received the B.S. degree in geographic Yanni Ma received the B.S. degree in geographical
information science, the M.E. degree in surveying information science in 2014 and the M.S. degree
and mapping engineering, and the Ph.D. degree in in surveying and mapping in 2018 from the China
surveying and mapping from the China University University of Geosciences, Beijing, China. She is
of Geosciences, Beijing, China, in 2015, 2018, and currently working toward the Ph.D. degree in sur-
2022, respectively. veying and mapping with the China University of
He is currently an Engineer with the China Aero Geosciences.
Geophysical Survey and Remote Sensing Center for She is currently an Engineer with the China Aero
Natural Resources, Beijing, China. His research in- Geophysical Survey and Remote Sensing Center for
terests include high spatial resolution remote sensing Natural Resources. Her research interests include
image processing and analysis, landslide information extraction, and semantic remote-sensing image analysis and deep-learning
understanding of urban functional zone. application.
1018 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 16, 2023

Yan Li received the B.S. degree in geographical Yueqin Zhu received the Ph.D. degree in cartography
information science from the China University of from the Technical University of Munich, Munich,
Geosciences, Beijing, China, in 2020. She is currently Germany, in 2012.
working toward the Ph.D. degree in surveying and She is currently a Senior Engineer with the Na-
mapping with the China University of Geosciences, tional Institute of Natural Hazards, Ministry of Emer-
Beijing, China. gency Management, Beijing, China. Her research
Her research interests include high-resolution re- interests include cloud computing, computational in-
mote sensing image procession and landslide risk telligence, and cartography.
evaluation.

Xiao Ling received the B.S. degree in geomatic

engineering from the Nanjing Normal University,
Nanjing, China, in 2017, and the M.E. degree in
surveying and mapping engineering from the China
University of Geosciences, Beijing, China, in 2021.
She is currently working toward the Ph.D. degree in
surveying and mapping with the China University of
Geosciences.
Her research interests include landslide hazard,
susceptibility analysis, and high-resolution remote
sensing image procession.