Computers & Geosciences 138 (2020) 104445

Contents lists available at ScienceDirect

Computers and Geosciences

journal homepage: www.elsevier.com/locate/cageo

Comparative study of landslide susceptibility mapping with different

recurrent neural networks
Yi Wang a, *, Zhice Fang a, Mao Wang a, Ling Peng b, Haoyuan Hong c, d, e, f, **
a
Institute of Geophysics and Geomatics, China University of Geosciences, Wuhan, 430074, China
b
China Institute of Geo-Environment Monitoring, Beijing, 100081, China
c
Department of Geography and Regional Research, University of Vienna, Vienna, 1010, Austria
d
Key Laboratory of Virtual Geographic Environment (Nanjing Normal University), Ministry of Education, Nanjing, 210023, China
e
State Key Laboratory Cultivation Base of Geographical Environment Evolution (Jiangsu Province), Nanjing, 210023, China
f
Jiangsu Center for Collaborative Innovation in Geographic Information Resource Development and Application, Nanjing, Jiangsu, 210023, China

A R T I C L E I N F O A B S T R A C T

Keywords: This paper aims to use recurrent neural networks (RNNs) to perform landslide susceptibility mapping in Yongxin
Landslides susceptibility mapping County, China. The two main contributions of this study are summarized as follows. First, the regular RNN is
Recurrent neural networks compared to its three variants in the case study of landslide susceptibility mapping for the first time, including
Long short term memory
long short term memory, gated recurrent unit and simple recurrent unit. Second, a sequential data representation
Gated recurrent unit
method is proposed to fully explore the predicting potential of RNNs. The study area consists of 364 historical
Simple recurrent unit
landslide locations that were divided into two parts: 255 (70%) for training and 109 (30%) for validation, and 16
landslide influencing factors were considered for spatial prediction. To validate the effectiveness of these RNN-
related methods, several objective measures of accuracy, recall, F-measure, Matthews correlation coefficient and
the receiver operating characteristic were used for evaluation. Experimental results demonstrate that very high
and high susceptible areas are concentrated in the northwest and south of Yongxin County, while landslides in
the central area are less prone to occur. Based on quantitative results, all the RNN-related methods achieved area
under the curve values above 0.83 and produced accurate prediction results with the optimized parameters.
Therefore, the RNN framework can be used as a useful tool for the landslide susceptibility mapping task to
mitigate and manage landslides.

1. Introduction qualitative and quantitative. Qualitative methods depend on the opin­

ions and judgments of experts, including geomorphological mapping
Landslides are dangerous geological disasters that pose a serious (Parry, 2011), heuristic or indexing methods (Pourghasemi et al.,
hazard to people’s lives and property. Nearly 70% of China’s areas are 2012b; Yalcin, 2008) and landslide inventory analysis (Galli et al., 2008;
mountainous, which provides innate conditions for the occurrence of Guzzetti et al., 2012), while quantitative methods conduct mathematical
landslides, resulting in a large number of landslide events (Nohani et al., analysis and establish a probability statistical model to analyse the
2019; Pham, 2018; Pham et al., 2019a). Although it is impossible to relationship between landslide occurrence and influencing factors. In
prevent landslide occurrence, disasters can be predicted and remedied particular, the quantitative methods are mainly deterministic and sta­
using appropriate methods and analysis. Among various landslide pre­ tistical (Tsangaratos et al., 2017). Specifically, deterministic methods
diction methods, landslide susceptibility mapping (LSM) is an effective construct protective components of unstable slopes (Clerici et al., 2006;
land use management technique, which can provide favourable support Go€kceoglu and Aksoy, 1996), while statistical methods employ multi­
for land managers’ decision-making (Hong et al., 2018b; Nourani et al., variate or binary statistical methods for evaluation (Regmi et al., 2014),
2014). such as logistic regression (Ayalew and Yamagishi, 2005), weight of
The LSM methods have been mainly divided into two groups, i.e., evidence (Neuha €user and Geomorphology, 2007), support vector

* Corresponding author. Institute of Geophysics and Geomatics, China University of Geosciences, Wuhan, 430074, China.
** Corresponding author. Department of Geography and Regional Research, University of Vienna, Vienna, 1010, Austria.
E-mail addresses: [email protected] (Y. Wang), [email protected], [email protected], [email protected] (H. Hong).

https://doi.org/10.1016/j.cageo.2020.104445
Received 10 May 2019; Received in revised form 4 February 2020; Accepted 12 February 2020
Available online 15 February 2020
0098-3004/© 2020 Elsevier Ltd. All rights reserved.
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Fig. 1. Location of the study area.

machine (Pham et al., 2019b), naïve Bayes (Pham et al., 2017a) and variants are connected using hidden layers that use a "state vector"
frequency ratio (Lee and Pradhan, 2007). pattern to obtain the previous important information for the next state.
In recent years, deep learning has received great attention in Second, it is a comparative study of the regular RNN and its variants for
different fields, such as speech recognition (Noda et al., 2015), auto­ landslide spatial prediction in the case study of Yongxin County, China.
matic translation (Pan et al., 2016), automatic driving (Hoermann et al., To verify the effectiveness of different methods, a series of statistical
2018), etc. The concept of deep learning stems from artificial neural measures were used, including accuracy (ACC), Matthews correlation
networks, which are multi-layer perceptron structures with multiple coefficient (MCC), recall, F-measure, receiver operating characteristic
hidden layers (Hinton and Salakhutdinov, 2006). More recently, deep (ROC) curve and area under the curve (AUC). In addition, the significant
learning techniques have been applied to natural disaster susceptibility difference between the proposed RNNs was validated by using a chi-
mapping tasks (Bui et al., 2019; Wang et al., 2019c; Xiao et al., 2018; square test.
Zhang et al., 2019). For example, we first used convolutional neural
networks in LSM and achieved promising results (Wang et al., 2019b). 2. Study area
As for the another effective deep learning technique of recurrent neural
networks (RNNs), it have been widely used in natural language pro­ 2.1. Description
cessing because they can capture dynamic information in data by peri­
odically connecting hidden layer nodes. Furthermore, the RNNs can Yongxin County has a total area of 2187 km2, located in the western
demonstrate excellent classification capability by making full use of of Jiangxi Province. This study area has a subtropical humid monsoon
context information (i.e., domain information). However, the main climate with sufficient sunshine, abundant rainfall and mild climate.
disadvantage of the regular RNN model is that it only considers the state The average annual rainfall is 1530.7 mm and the rainy season is from
of the most recent moment. To solve this problem, several variants of March to August. The geological environment is part of the South China
RNN have been presented by introducing the concept of cell state, fold system, and the structural changes are very obvious. Also, folds and
including long short term memory (LSTM) (Chen et al., 2015; Graves, faults are well developed. Excepting for the Sinian, Silurian and Tertiary,
2012), gated recurrent unit (GRU) (Cho et al., 2014) and simple recur­ the Cambrian to Quaternary strata were exposed to the study area with a
rent unit (SRU) (Lei et al., 2018). It should be noted that Mutlu et al. total thickness of over 20,000 m.
(2019) performed landslide susceptibility analysis using RNN, demon­ Historical landslide locations provide the underlying conditions for
strating its effective prediction ability. Moreover, the RNN technique has predicting the likelihood of future landslides (Wang et al., 2016). As
been used in landslide detection as well. For example, the RNN model shown in Fig. 1, local government departments in Jiangxi Province
was used for landslide deformation prediction (Chen et al., 2015). provided a total of 364 landslide locations. According to statistics, 70%
Moreover, Xu et al. successfully predicted landslide displacement of these landslides are rotational landslide and 30% of them belong to
induced by rainfall using the LSTM model (Xu and Niu, 2018). However, translational landslide. The largest and smallest landslides being 750,
these studies mainly focused on the assessment of monomer landslide 000 m2 and 32 m2, respectively. All the landslides in the study area can
risk. be divided into three scales: large (>1000 m2), moderate (400–1000 m2)
To explore more excellent deep learning techniques and further and small (<400 m2). According to the local government report, almost
extend the application of the RNN technique to regional-scale landslide all landslides occurred during or after heavy rain, of which 64.3% of the
susceptibility prediction, we present in this study an RNN framework for landslides occurred with 100 mm of precipitation every day. Moreover,
LSM. The two main contributions of this study are summarized as fol­ the lives of about 2174 people were affected by these landslide disasters,
lows. First, a sequential data representation method is proposed to fully and their property suffered huge losses. Therefore, it is necessary to
explore the predicting potential of RNNs. The regular RNN and its perform LSM in this area to prevent and mitigate the adverse effects of

2
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Fig. 2. Field photos of some typical landslides.

2019). In this study, 16 landslide influencing factors were considered,

Table 1
including lithology, distance to faults, land use, soil, distance to roads,
Lithological group of the study area.
altitude, slope, aspect, profile curvature, plan curvature, stream power
Group Unit name Lithology index (SPI), topographic wetness index (TWI), sediment transport index
A Lianhe group, Tangbian group, Conglomerate, mudstone (STI), rainfall, distance to rivers and normalized difference vegetation
Hekou group index (NDVI). Lithology is a key factor that has been used in previous
B Zhangzong group, Zhongpeng Shale studies (Nsengiyumva et al., 2019; Shirzadi et al., 2018b; Yang et al.,
group, Yunshan group
C Changlong group, Oujia chong Limestone and sandy shale
2019). Table 1 lists the description of each group in the study area
group, (Wang et al., 2019a), which is provided by China Geology Survey (htt
D Huamian gong group, Shi kou Limestone and siliceous slate p://www.cgs.gov.cn). It affects the development of topographical
group structure and leads to differences in rock strength. Moreover, the
E Duier shi group, Jueshan gou group Slate, black carbonaceous siliceous
geological structure has a critical influence on the stability of the slope,
slate
F Longtang group, Qi baoshan group, Coal seam, cherty limestone and so this study chose the factor of distance to fault (Basu and Pal, 2018).
Chang xing group siliceous rocks Landslide occurrence is related to land use since it reflects the human
G Ba cun group, Liu jiaohe group Carbonaceous slate activity and variation of land cover (Nicu and Asa �ndulesei, 2018; Z^ezere
H Gu feng group, Qixia group, Xiao Carbonaceous shale; cherty et al., 2017). The soil structure is loose and easy to accumulate water,
Jiangbian group limestone, siliceous rocks
I Huang xie group, Hai hui group, Xi Two long granite
forming a large number of slope bodies with sufficient sliding space (Bui
hua group et al., 2017). The factor of distance to roads can reflect the intensity of
J Chang lejie group, Gu poshan ranite human influence (Peshevski et al., 2019; Steger et al., 2016). Altitude
group, San jiangkou Group directly influences different environmental conditions, i.e., tectonism
K Zishan group, Yang jiayuan group Dolomite and coal seam
and climate (Zhu et al., 2018, 2019). Slope determines the steepness of
L Hu tian group Biolimestone
M Dui ershi group, Shi kou Group Slate; green slate, carbonaceous the terrain surface and the stress degree of landslides (Zhou et al., 2018).
slate Aspect determines the received solar radiation degree, which can in­
N Xia shan group, Yi jiawang group, Sandstone, shale, dolomite fluence the soil moisture and slope stability (Dou et al., 2019c; Nepal
Qi ziqiao group et al., 2019). Plane curvature affects the convergence and dispersion of
O Fu fang group, Taihe group, Tang Two long granite
hu group
the flow, whereas profile curvature influences the acceleration and
P Ba cun group, Shui shi group Slate, sandstone deceleration of the flow (Oh and Pradhan, 2011). SPI measures surface
Q Ba cun group, Gao tang group Green sandstone, silty slate runoff erosion capability, and TWI predicts areas susceptible to the
surface of saturated soil (Regmi et al., 2010). STI describes topographic
variables of water and sediment transport in landslides (Pourghasemi
landslides. In this study, field surveys, google earth image interpretation
et al., 2012a). Rainfall may cause a large amount of rainwater to infil­
and historical landslide records were combined to evaluate the likeli­
trate, causing the soil layer on the slope to saturate, which increases the
hood of landslides in a region and to compile a landslide inventory map
weight of the sliding body and causes landslides (Pham et al., 2016; Yu
(Crozier, 2018). The field photos of some typical landslides are shown in
et al., 2016). Rivers is another important factor that affects the soil
Fig. 2.
saturation, and erosive undercutting of rivers can also influence the slide
surface (Erener et al., 2017). NDVI represents vegetation conditions and
2.2. Landslide Geodatabase groundwater content, which may affect the development of landslides
(Ding et al., 2017). It should be noted that a digital elevation model
The selection of landslide influencing factors has an important (DEM) for the study area was obtained from the ASTER GDEM website
impact on the final LSM (Meusburger and Alewell, 2009; Ozer et al.,

3
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Table 2 Chen et al., 2018):

Descriptions of landslide influencing factors. 8
>
> jrj � 0:3 no linear correlation
Data type Factors Source Description <
0:3 < jrj � 0:5 low linear correlation
Geologic lithology China Some properties that reflect rock > 0:5 < jrj � 0:7
> significant linear correlation
:
Geology characteristics jrj � 0:7 highly linear correlation
Survey
distance to Soil and The effect of significant relative
fault Geology displacement on the lateral block 3.2. Multi-collinearity analysis
Maps
Land-related Land use Landsat Root reinforcement of soil, Multi-collinearity refers to inaccurate regression estimates due to the
ETMþ surface runoff regulation high correlation between variables in multiple regression models
soil Soil and Different soils have different
(O’brien, 2007). In this analysis, two indicators of tolerance (TOL) and
Geology characteristics
Maps variance inflation factor (VIF), were used for evaluation. Let X ¼
distance to GIS Database Road impact on landslide fX1; X2; X3 ; :::; Xn; g present the independent variable data set and R2j
road
denote the complex measurement the coefficient for the regression of
Topographic altitude DEM A little bit of vertical lead in the
space to a certain level other independent variables when the independent variable Xj is
slope DEM Overland and sub-surface flow regressed on all other predictor variables (Wang et al., 2017). The VIF
velocity value is calculated as follows:
aspect DEM Different slope directions have .� �
different solar radiation
VIF ¼ 1 1 R2j (2)
intensities
profile DEM Affects the acceleration and
curvature deceleration of the flow The TOL is numerically the reciprocal of VIF. Specifically, if R2j ¼ 0,
plan DEM Affects the convergence and then VIF ¼ 1 and TOL ¼ 1, which means that there is no linear corre­
curvature dispersion of flows
lation between Xj and the others; if the value of R2j is very close to 1,
Water- SPI DEM Erosive power of water flow
related TWI DEM Predict areas susceptible to VIF→ þ ∞ and TOL ¼ 0, which represents highly correlation between
surface effects of saturated land them. In particular, if VIF < 5(TOL > 0:2) or VIF < 10 (TOL > 0:1), the
STI DEM Moisture and sediment transport variables are not collinear and are independent of each other (O’brien,
in specific objects
2007).
rainfall GIS Database Lead to instability in the soil
distance to GIS Database Erosion of river water
river 3.3. Information gain
NDVI Landsat Reflecting the difference in
ETMþ ground moisture content Information Gain is a landslide influencing factor feature selection
model for LSM (Tien Bui et al., 2016). It can determine the weight of
(http://gdem.ersdac.jspacesystems.or.jp). The detailed descriptions of influencing factors and effectively identify the importance of factors for
these factors are shown in Table 2. All the landslide influencing factors landslide occurrence. The information gain is determined by calculating
were converted into a raster form with a spatial resolution of 25 m. The the entropy reduction of the output category y to which the input factor
thematic maps of these factors are shown in Fig. 3. xi corresponds.
IGðy; xi Þ ¼ EðyÞ Eðyjxi Þ (3)
3. Material and methodology
where Eðyjxi Þ is the conditional entropy and EðyÞ is a priori Shannon
Fig. 4 illustrates the flowchart of the RNN framework. The five main entropy, and they are calculated as follows:
steps of this framework is summarized as follows. First, landslide
X
n
influencing factors are selected for modelling. Then, these factors are EðyÞ ¼ yi log2 ðyi Þ (4)
analyzed and screened based on their importance. Next, the landslide i¼1

data are converted to sequential representation, and training and vali­

X
n
dation sets are obtained by randomly dividing the available landslide Eðyjxi Þ ¼ yi EðyÞ (5)
historical locations. In the following, the RNN model and its variants are i¼1
constructed to obtain landslide susceptibility maps. Finally, the resultant
Average merit (AM) indicates the relationship between influencing
maps are evaluated and compared using different statistical measures.
factors and landslide occurrence. The higher the weight, the greater the
contribution of the corresponding factors to occurrence of landslides. If
3.1. Correlation analysis the AM value is less than or equal to 0, then the influencing factor is
considered irrelevant and should be rejected.
The Spearman’s rank correlation coefficient is a statistical indicator
that reflects the closeness of correlation between two variables X and Y 3.4. Certainty factor
(Gao, 2011). The range of the coefficient is between 1 and 1, repre­
senting the total negative linear and positive linear correlation, The certainty factor (CF) method can be employed to evaluate the
respectively. In particular, 0 means no linear correlation. Given a pair of correlation between landslide occurrence and influencing factors. This
random variables X and Y, the correlation coefficient can be calculated method is capable of handling the problem of the combination of
as follows: different data layers, the heterogeneity and uncertainty of the input
covðX; YÞ data. The CF value can be calculated as follows (Chung and Fabbri,
rðX; YÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffipffiffiffiffiffiffiffiffiffiffiffiffiffi (1) 1993; Oh and Pradhan, 2011):
Var½X� Var½Y�

where cov(X, Y) is the covariance of X and Y, Var[X] and Var[Y] are the
variances of X and Y, respectively. In practice, the degree of linear
correlation is summarized as follows (Booth et al., 1994; Bui et al., 2016;

4
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Fig. 3. Landslide influencing factor maps. (a) Lithology, (b) distance to faults, (c) land use, (d) soil, (e) distance to road, (f) altitude, (g) slope, (h) aspect, (i) profile
curvature, (j) plan curvature, (k) stream power index (SPI), (l) topographic wetness index (TWI), (m) sediment transport index (STI), (n) rainfall, (o) distance to river
and (p) normalized difference vegetation index (NDVI).

5
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Fig. 3. (continued).

8 pp
> a pps CF value ranges between 1 and 1. A positive value indicates a high
< ppa ð1 pps Þ ppa � pps
>
probability of landslides, where as a negative value indicates a low
CF ¼ ​ (6)
>
> ppa pps probability of landslides. In particular, if this value is equal to 0, it is
: ppa < pps
pps ð1 ppa Þ difficult to determine the certainty between influencing factors and the
occurrence of landslides.
where ppa is the conditional probability of a landslide in a certain class a,
and pps is the prior probability of a landslide in the entire study area. The

6
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Fig. 4. Flowchart of the recurrent neural network (RNN) framework.

Fig. 5. Typical RNN architecture.

3.5. Recurrent neural network and its variants Fig. 6. Long short term memory (LSTM) architecture.

3.5.1. Recurrent neural network

RNN can capture dynamic information in serialized data by period­ ht ¼ σ ðWh xt þ Uh ht 1 þ bh Þ (7)
ically connecting hidden layer nodes, as shown in Fig. 5. Unlike other �
neural networks, it can take advantage of sequential information. Spe­ yt ¼ σ Wy ht þ by (8)
cifically, in traditional neural networks, all inputs (and outputs) are
considered independent. In the RNN method, each unit is associated where σ (⋅) is the loss function of the training sample sequence, W and U
with other units in the hidden layer at different time intervals. Thus, are parameter matrices, b is the corresponding bias vector.
information can be passed from one layer to the next in the network (Xu
and Niu, 2018). 3.5.2. Long short term memory network
They are shared throughout the RNN network, which exactly LSTM is a special RNN structure that learns long-term dependency
embody the idea of "loop feedback". The simple RNN is always imple­ information. It was proposed by Hochreiter and Schmidhuber (Chen
mented using Elman or Jordan network. Let xt, ht, and yt be the input et al., 2015), and later improved by Graves (2012). Unlike a single
vector, the hidden state vector and the output vector at time t, respec­ neural network layer, LSTM carries out information protection and
tively, we can obtain control through three structures of input gate, forget gate and output
gate, which interact in a very special way. Specifically, the input gate

7
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Fig. 7. Gated recurrent unit (GRU) architecture. Fig. 8. Simple recurrent unit (SRU) architecture.

controls how much new information is added, the forget gate decides whereσg and σ h represent the sigmoid function and the hyperbolic
what information should be discarded from the state of the previous tangent, respectively, W and U are parameter matrices, and b is the
unit, and the output gate modulates the extent to which the state of the corresponding bias vector.
unit is filtered (Donahue et al., 2015). Fig. 6 demonstrates the LSTM
architecture. Let ft, It, and Ot be the activation vectors of the forget gate, 3.5.4. Simple recurrent unit
input gate and output gate at time t, respectively, and Ct cell state vector Lei et al. (2018) proposed a SRU model based on the LSTM and GRU
at time t, they can be calculated as follows: models. Similar to LSTM and GRU, SRU utilizes the gate structure to
� control the transmission of information flow. The main design principle
ft ¼ σ Wf1 xt þ Wf2 ht 1 þ bf (9)
of SRU is that the calculation of the gate depends only on the loop of the
current input. Therefore, the calculation of the model only multiplying
It ¼ σðWi1 xt þ Wi2 ht þ bi Þ (10)
point-by-point matrix dependent on the previous time step, which
1

Ot ¼ σðWo1 xt þ Wo2 ht þ bo Þ (11) makes the network easy to parallelize. The SRU sets the forget gate and
1
reset gate, so that the SRU is a deeper network because each layer re­
Ct ¼ ft *Ct þ It *C
~t (12) quires less computational power and a higher processing speed. Fig. 8
1
demonstrates a typical SRU architecture. Let ft, rt, It and Ot be the acti­
where W1 represents the weight between the input node and the hidden vation vectors of the forget gate, reset gate, input gate and output gate at
node, W2 represents the weight connecting the hidden node to the time t, respectively, they can be calculated as follows:
output node, and b is the corresponding bias vector. �
ft ¼ σ Wf xt þ Uf � ht 1 þ bf (16)

3.5.3. Gated recurrent unit It ¼ ft � ht þ ð1 ft Þ � ðWxt Þ (17)

As stated in the publication of (Cho et al., 2014), GRU is a variant of
1

LSTM because the number of its parameters is less than those of LSTM, rt ¼ σðWr xt þ Ur � ht 1 þ br Þ (18)
so the final model is simpler than the regular LSTM model. GRUs have
been very popular due to their better performance on smaller data sets Ot ¼ rt � ht þ ð1 rt Þ � x t (19)
(Chung et al., 2014b). The GRU merges the input and forget gates into a
single update gate with the addition of a reset gate. The update gate where � is the point-wise multiplication operation, W and U are
controls the degree to which the status information of the previous parameter matrices, b is the corresponding bias vector. (Jiang et al.,
moment is brought to the current state. The larger the value of the up­ 2018).
date gate, the more the status information is brought in at the previous
moment. The reset gate controls the degree of ignoring the status in­
3.6. Data representation for recurrent neural networks
formation of the previous moment. The smaller the value of the reset
gate, the more it is neglected. Fig. 7 demonstrates a typical GRU ar­
From the mathematical background of the above four RNN-based
chitecture. Let Zt, rt, and Ot be the activation vectors of the update gate,
methods, it can be concluded that RNN is good at processing sequen­
reset gate and output gate at time t, respectively, they can be calculated
tial inputs through special recurrent hidden states. Therefore, proper
as follows:
data representation is critical to exploring the predictive potential of
Zt ¼ σ g ðWz xt þ Uz ht 1 þ bz Þ (13) RNNs. In this section, the landslide data representation for RNNs is
presented, as shown in Fig. 9. Firstly, each landslide influencing factor
rt ¼ σg ðWr xt þ Ur ht 1 þ br Þ (14) can be viewed as a single-band image, and all factor layers are stacked
together. Then, as described in section 3.3, the importance of the
Ot ¼ ð1 Zt Þ * ht 1 þ Zt *σ h ðWh xt þ Uh ðrt ht 1 Þ þ bh Þ (15) influencing factors is calculated by using an information gain algorithm.
Next, these factor layers are sorted in descending order of importance.

8
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Fig. 9. Data representation for RNNs.

Thus, each pixel can be converted to a sequential sample from the

Table 3
perspective of importance. In this way, the most important factors are
Multi-collinearity analysis for influencing factors.
sent to the RNN structure first, and the least important factors are sent to
the model last. In addition, due to the recurrent structure of the RNN, the Variable Collinearity statistics
most critical information that contributes to landslide occurrence will be TOL VIF
retained and passes to the next hidden state, which is useful for the final altitude 0.456 2.193
landslide susceptibility prediction. aspect 0.967 1.034
fault 0.910 1.098
Land use 0.516 1.938
3.7. Model evaluation methods lithology 0.751 1.332
NDVI 0.637 1.569
The verification of the pros and cons of the model is the key condition plan curvature 0.675 1.481
profile curvature 0.785 1.274
for the landslide to occur (Wang et al., 2019a). It is very common to use
rainfall 0.862 1.161
ROC for analysis and evaluation. The ROC curve is an indicator of distance to river 0.825 1.212
continuous variables of data specificity and sensitivity. The AUC rep­ distance to road 0.925 1.082
resents the predicted value of the model. The larger the area under the slope 0.350 2.854
curve, the better the classification (Tsangaratos et al., 2017). At the soil 0.806 1.240
SPI 0.653 1.532
same time, the prediction ability of the landslide model was evaluated STI 0.369 2.707
by using statistical indicators such as ACC, recall, F-measure, Matthews TWI 0.378 2.649
correlation coefficient (MCC), and these statistic calculations are as
follows (Ercanoglu and Gokceoglu, 2002; Matthews, 1975; Powers,
2011): where TP and TN represent the correct classification of landslides and
TP þ TN non-landslides, and FP and FN denote misclassified landslides and non-
ACC ¼ (20) landslides, respectively. For accuracy, recall and F-measure, the value is
TP þ FP þ TN þ FN
between 0 and 1. The higher the value, the better the model. For MCC, it
TP is between 1 and þ 1. Specifically, þ1 for perfect predictions, 0 for
Recall ¼ (21)
TP þ FN random predictions and 1 for complete inconsistency between pre­
dictions and observations.
F measure ¼
2 � TP
(22) In addition, the chi-square test has become a common tool for vali­
2 � TP þ FP þ FN dating statistical differences between two prediction methods (Tallarida
and Murray, 1987). Based on a priori assumption, there is no statistical
TP � TN FP � FN
MCC ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (23) difference between the landslide methods, and then the chi-square value
ðTP þ FPÞ � ðTP þ FNÞ � ðTN þ FPÞ � ðTN þ FNÞ
is calculated, and the difference value of the method is used to quantify

9
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

alt - altitude, asp - aspect, fau - distance to fault, lan - land use, lit - lithology, ndv – NDVI, pla - plan curvature, pro - profile curvature, rai – rainfall, riv-distance to river, roa - distance to road, soi – soil, spi – SPI, sti – STI, twi
0.278
0.007

0.118
0.166
0.451

0.029
0.111
0.062
0.468
0.084
0.011
0.276

0.345

0.336
0.208
1.000
twi

0.009
0.015
0.288

0.330
0.333

0.161
0.215

0.176
0.059
0.099
0.085
0.524
0.064
0.512
1.000
sti

0.006
0.009
0.089

0.152
0.121

0.058
0.083

0.080
0.034
0.015
0.032
0.094
0.002
1.000
spi

0.002

0.299

0.013
0.021
0.092

0.020

0.116
0.200
0.009

0.065
0.052
0.128
1.000
soi

0.005
0.018
0.490

0.075
0.545

0.246
0.328
0.035

0.081
0.185
0.133
1.000
slo

Fig. 10. Average merit (AM) of each landslide influencing factor.

0.188

0.020
0.164
0.005
0.034

0.030
0.126
0.006

0.115
0.169
1.000
roa

the results (Bagui, 2005; Pham et al., 2017b). If the chi-square value
exceeds 3.841 and the p-value is less than 0.005, there is a significant
difference between the two methods.
0.260

0.046
0.319
0.059
0.002

0.016
0.126
0.009

0.035
1.000
riv

4. Results

4.1. Selection of landslide influencing factors

0.001
0.019
0.073

0.021

0.007
0.141

0.140

0.003

1.000
rai

Table 3 lists the collinearity analysis between those influencing

factors under the condition of 95% confidence and Table 4 lists a
Spearman correlation matrix of 120 pairs of landslide influencing fac­
0.105

0.027
0.044
0.421
0.002
0.000
0.081

1.000

tors. It can be observed that all correlation coefficients are lower than
pro

0.7 and the VIF values of those factors are lower than 5 (TOL > 0:2).
Therefore, all the factors are independent of each other and there is no
0.039

collinearity between them.

0.050
0.004
0.000

0.010
0.013
1.000

Fig. 10 demonstrates the factor weight of each factor determined by

pla
Spearman correlation between landslide influencing factors in conditional independence test.

information gain. Based on these results, the lithology factor has the
highest AM value of 0.2573, indicating that it is the dominant factor to
0.028
0.044
0.471

induce landslide occurrence. The AM values of altitude, NDVI, soil,

0.444

0.254
1.000
ndv

aspect, and land use are all between 0.1 and 0.2, whereas the AM values
of the other factors are between 0 and 0.1, indicating that these factors
can contribute to landslide occurrence.
0.002
0.242
0.294
0.369

1.000

4.2. Analysis of landslide influencing factors

lit

To investigate the relationship between landslides and all the influ­

0.635

encing factors, the CF model was used. Table 5 lists reclassification re­
0.022
0.011
1.000
lan

sults of each factor and the corresponding CF values. For elevation, it

shows a positive correlation for the 250–400, 400–700, 700–1000
classes and a negative correlation for the remaining classes. For aspect,
0.050

the flat, northeast, east and southwest classes exhibit positive correla­
0.001
1.000
fau

tion and the remaining classes are negatively correlated. For distance to
fault, the CF value gradually decreases as the distance is increased. For
land use, the grass and bare classes are positively correlated to land­
0.009
1.000

slides. For lithology, there are 7 classes showing a positive correlation,

asp

the O and P classes are more prone to high landslide susceptibility. For
NDVI, the 0.2–0.3 and 0.3–0.4 classes demonstrate a very strong positive
correlation. For rainfall, the highest CF value of the >1100 class in­
1.000

dicates that the amount of precipitation greatly affects landslide

alt

occurrence. For slope, a negative correlation can be seen when it is

larger than 60� and the maximum CF values are 0.85 at 50� –60� . For
Variable

soil, the ACu, ALh and ACh classes are more likely to cause landslides.
Table 4

- TWI.
ndv

pro
asp

roa

twi
fau
lan

pla

spi
soi
slo
riv
rai
alt

sti
lit

10
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Table 5
Analysis of the relationship between each impact factor and landslide occurrence.
Influencing factor Class No. of pixels in domain Percent of domain (%) No. of landslide Percent of landslide (%) CF

Altitude (m) <150 876,142 24.84 4 1.1 0.96

150–250 947,989 26.88 55 15.11 0.47
250–400 822,620 23.33 127 34.89 0.37
400–700 663,718 18.82 146 40.11 0.60
700–1000 180,954 5.13 30 8.24 0.43
>1000 35,309 1 2 0.55 0.48
Slope aspect Flat 26,229 0.74 13 3.57 0.89
North 460,854 13.07 16 4.4 0.69
Northeast 409,633 11.62 53 14.56 0.23
East 472,022 13.38 90 24.73 0.52
Southeast 458,681 13.01 113 31.04 0.66
South 466,466 13.23 48 13.19 0.00
Southwest 381,226 10.81 23 6.32 0.45
West 422,817 11.99 8 2.2 0.83
Northwest 428,804 12.16 0 0 1.00
Distance to fault (m) 0–2000 1,605,946 45.54 191 52.47 0.15
2000–4000 1,030,730 29.23 110 30.22 0.04
4000–6000 626,737 17.77 55 15.11 0.17
6000–8000 234,027 6.64 8 2.2 0.70
>8000 29,292 0.83 0 0 1.00
Land use water 42,077 1.19 0 0 1.00
forest 1,231,417 34.92 103 28.3 0.21
grass 926,977 26.28 213 58.52 0.62
farmland 468,788 13.29 12 3.3 0.77
bare 150,316 4.26 30 8.24 0.55
residential 707,157 20.05 6 1.65 0.93
Lithology Group A 784,838 22.25 3 0.82 0.97
Group B 429,108 12.17 62 17.03 0.32
Group C 2,311 0.07 0 0 1.00
Group D 233,843 6.63 39 10.71 0.43
Group E 77,394 2.19 6 1.65 0.27
Group F 6,728 0.19 0 0 1.00
Group G 189,655 5.38 18 4.95 0.09
Group H 145,633 4.13 0 0 1.00
Group I 17,333 0.49 0 0 1.00
Group J 5 0 0 0 1.00
Group K 423,127 12 14 3.85 0.71
Group L 9,764 0.28 0 0 1.00
Group M 234,472 6.65 46 12.64 0.54
Group N 446,643 12.66 67 18.41 0.35
Group O 48,645 1.38 36 9.89 0.97
Group P 4,670 0.13 2 0.55 0.86
Group Q 472,563 13.4 71 19.51 0.35
NDVI <0.1 1,821,056 51.64 60 16.48 0.71
0.1-0.2 860,462 24.4 74 20.33 0.18
0.2-0.3 763,478 21.65 195 53.57 0.67
0.3-0.4 81,671 2.32 35 9.62 0.86
>0.4 65 0 0 0 1.00
Plan curvature <-0.25 1,242,908 35.24 164 45.05 0.25
0.75 1,369,415 38.83 98 26.92 0.33
>0.5 914,409 25.93 102 28.02 0.08
Profile curvature <-1 626,562 17.77 95 26.1 0.36
(-1)-0.2 1,387,942 39.35 91 25 0.39
>0.2 1,512,228 42.88 178 48.9 0.14
Rainfall <800 27,289 0.77 0 0 1.00
800–900 309,590 8.78 8 2.2 0.77
900–1000 1,015,248 28.79 113 31.04 0.08
1000–1100 1,057,419 29.98 83 22.8 0.26
>1100 1,117,186 31.68 160 43.96 0.32
Distance to river (m) <200 1,198,190 33.97 88 24.18 0.31
200–500 1,321,706 37.48 154 42.31 0.13
500–800 735,755 20.86 94 25.82 0.22
>800 271,081 7.69 28 7.69 0.00
Distance to road (m)- <300 737,514 20.91 45 12.36 0.44
300–1300 1,736,532 49.24 213 58.52 0.18
1300–2300 787,621 22.33 84 23.08 0.04
>2300 265,065 7.52 22 6.04 0.22
Slope (� ) 0–10 1,466,412 41.58 64 17.58 0.61
10–20 1,036,916 29.4 111 30.49 0.04
20–30 620,572 17.6 95 26.1 0.37
30–40 305,484 8.66 66 18.13 0.59
40–50 87,469 2.48 24 6.59 0.70
50–60 9,679 0.27 4 1.1 0.85
>60 200 0.01 0 0 1.00
(continued on next page)

11
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Table 5 (continued )
Influencing factor Class No. of pixels in domain Percent of domain (%) No. of landslide Percent of landslide (%) CF

Soil ATc 584,915 16.59 1 0.27 0.99

ACu 346,372 9.82 77 21.15 0.61
ALh 136,843 3.88 22 6.04 0.40
ACh 2,275,477 64.52 263 72.25 0.12
RGc 183,125 5.19 1 0.27 0.95
SPI <400 3,411,208 96.72 334 91.76 0.06
400–2000 99,870 2.83 22 6.04 0.60
2000–5000 12,610 0.36 8 2.2 0.95
5000–12000 2,675 0.08 0 0 1.00
>12000 369 0.01 0 0 1.00
STI <15 2,701,877 76.61 198 54.4 0.32
15–100 781,291 22.15 155 42.58 0.54
100–200 34,803 0.99 8 2.2 0.62
200–300 6,041 0.17 3 0.82 0.89
>300 2,720 0.08 0 0 1.00
TWI <5 1,129,914 32.04 159 43.68 0.30
5–7 1,678,597 47.6 141 38.74 0.21
7–9 546,447 15.49 40 10.99 0.32
9–11 150,377 4.26 19 5.22 0.21
>11 2,1397 0.61 5 1.37

the direction of northwest and southeast. In Fig. 11 (b), almost all the
Table 6
historical landslide locations are distributed in the very high susceptible
Parameter settings of RNNs for landslide susceptibility mapping.
zone. In Fig. 11 (c), the very high susceptible zone is distributed in the
Optimized Suitable Value Description direction of northwestern and the southern and it was less than those of
Parameter
RNN LSTM GRU SRU the other three models. In Fig. 11 (d), the very high susceptible zone is
Epoch 50 100 150 300 Train all concentrated in the direction of southwest and some areas without the
sample times historical landslide locations were classified into the high-prone zone.
Batch size 64 32 64 32 The size of Table 7 lists the percentage of landslide susceptibility classes of each
each batch of model. Landslide susceptibility assessment can produce a sensitive
data
Dropout 0 0 0.5 0 Mitigate the
model that focuses on highly susceptible areas and processes them in a
occurrence of concise manner (Kornejady et al., 2017a, 2017b). It can be observed that
over-fitting the four methods have the highest proportion of the very low susceptible
Optimizer Adam Adaptive class in the landslide susceptibility maps. Meanwhile, high susceptible
moment
classes occupy the lowest proportion in the resultant maps by RNN, GRU
estimation
Loss Categorical Binary and SRU, respectively, while low susceptible class takes up the lowest
function cross- classification proportion in LSTM-derived map. The evaluation of the practicability of
entropy problem the four models can be measured using the sum of the percentages of the
Learn rate 0.001 Determines the very high and high susceptible classes. From Table 7, the GRU model is
speed of weight
update
more practical than its variants due to the lowest practicality. It can be
also seen that the landslide density (LD) value gradually increases as the
landslide susceptible level is increased from very low to very high.
4.3. Models validation and comparison The results of the four models were assessed using the test set.
Table 8 lists the statistical results of four evaluation indicators of ACC,
All the influencing factors were input into the RNN, LSTM, GRU and recall, F-measure and MCC. It can be observed that GRU has the highest
SRU models, which were implemented in Python under the Tensorflow ACC of 0.7890, which is over 0.02 higher than that of SRU (0.7569). The
(http://www.tensorflow.org) and Scikit-learn (https://scikit-learn. highest recall, F-measure and MCC values were achieved by LSTM. In
org/stable) frameworks. To construct training and test sets, 364 non- addition, the performance of the SRU model is worse than that of the
landslide points were randomly selected from areas where no land­ other three models in terms of all the statistical measures.
slide occurred. Then, 364 landslide samples and 364 non-landslide Fig. 12 plots the ROC curves of the four models. It can be observed
samples were used to construct the training and validation sets: 70% that all the AUC values by the four models were above 0.83, indicating
of landslides (255) and non-landslides (255) for training and the that the RNN techniques can demonstrate very satisfactory prediction
remaining data (109 and 109) for testing. In addition, to objectively capability. Furthermore, the GRU model was better than the other
describe the behaviors of the four RNN-based models, five-fold cross- models due to its highest AUC value. Table 9 list the chi-square and p
validation was used to find the optimal parameters. Table 6 list the values between different models. It can be observed that all the chi-
parameter settings of the four models for experiments. The output re­ square values were much larger than 3.841 and all the p values were
sults were analyzed and processed by ArcGIS and the natural breakpoint lower than 0.005, thus there is a significant difference between the
method was used to classify a landslide susceptibility map into five landslide models.
groups, i.e., very high, high, moderate, low and very low. Finally, the Producing accurate and reliable susceptibility maps is important for
landslide susceptibility maps of the four models were obtained and preventing landslide disasters. Recently, a novel performance evalua­
shown in Fig. 11. It can be observed that the historical landslide loca­ tion approach using the photogrammetric technique was presented to
tions are basically in the high susceptibility area, and the low suscepti­ assess landslide susceptibility maps (Sevgen et al., 2019). It is instructive
bility areas are distributed in the northeast direction of the study area. In in verifying the reliability of the landslide susceptibility map. However,
the northwest and south of the study area, the landslide locations are we tried to collect the remote sensing images of the study area, but we
evenly distributed. In Fig. 11 (a), the proportion of the very high class is could not obtain the stereo pairs to complete the same experiment as in
very lower than that of the other three models, mainly concentrated in the literature. To cope with this situation, we adopted another strategy

12
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Fig. 11. Landslide susceptibility maps by (a) RNN, (b) LSTM, (c) GRU and (d) SRU.

Table 7
Landslide distribution in predicted landslide susceptible zones.
Model Landslide-susceptible zones No. of pixels in domain Area of zones (%) Practicality (IV þ V) No. of landslide Landslides percentage (%) LD

RNN Very low (I) 1,481,193 42.00 29.31% 22 6.04 0.14

Low (II) 549,066 15.57 29 7.97 0.51
Moderate (III) 462,686 13.12 46 12.64 0.96
High (IV) 457,976 12.99 84 23.08 1.78
Very high (V) 575,811 16.33 183 50.27 3.08
LSTM Very low (I) 1,156,771 32.80 37.38% 3 0.82 0.03
Low (II) 469,297 13.31 12 3.30 0.25
Moderate (III) 582,284 16.51 56 15.38 0.93
High (IV) 619,580 17.57 105 28.85 1.64
Very high (V) 698,800 19.81 188 51.65 2.61
GRU Very low (I) 1,740,859 49.36 25.91% 27 7.42 0.15
Low (II) 433,789 12.30 31 8.52 0.69
Moderate (III) 438255 12.43 63 17.31 1.39
High (IV) 455,477 12.91 98 26.92 2.08
Very high (V) 458,352 13.00 145 39.84 3.07
SRU Very low (I) 1,259,717 35.72 30.84% 30 8.24 0.23
Low (II) 618,008 17.52 32 8.79 0.50
Moderate (III) 561,188 15.91 60 16.48 1.04
High (IV) 543,742 15.42 81 22.25 1.44
Very high (V) 544,077 15.43 161 44.23 2.87

to assess the resultant susceptibility maps based on the available re­

Table 8 sources and data. The purpose of this strategy is to verify whether the
Performance of the four models. predicted high susceptible areas are likely to have a landslide hazard
Statistical measures RNN LSTM GRU SRU using two GaoFen-1 high-resolution satellite images with a spatial res­
ACC 0.7615 0.7660 0.7890 0.7569 olution of 8 m, acquired on December 5, 2013 and September 28, 2019,
Recall 0.8624 0.8633 0.8073 0.7928 respectively. To better show the observations, three sub-regions of the
F-measure 0.7833 0.8050 0.7928 0.7745 study area that have never experienced a landslide in the past were
MCC 0.5339 0.5791 0.5784 0.5201
selected for evaluation. Fig. 13 shows satellite images and susceptibility
maps for the three sub-regions. It can be observed that there were no
landslides in the three sub-regions in 2013. For the sub-region (a), we
can see from the satellite image of 2019 that a landslide event has

13
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

determines the gradient value and the frequency of weight updating. To

obtain its optimal parameter, the batch sizes are ranged from 25 to 28. In
Fig. 14 (b), the highest AUC values of RNN and GRU were achieved
when the batch size was fixed to 25, whereas LSTM and SRU obtained
the highest AUC values with the batch size of 26. The parameter of
dropout can theoretically prevent the over-fitting problem. Specifically,
it randomly inactivates some hidden neurons in the neural network. In
Fig. 14 (c), the highest AUC values of the GRU were achieved when
dropout was fixed to 0.5. The performance of the gradient descent
method is closely related to the setting of the learning rate, thus it is
necessary to set its value in a proper range. Commonly used optimizers
are Adam, RMSprop, Nadam, Adamelta, and Adamax. In Fig. 14 (d), the
highest AUC values of the four models were achieved when the Adam
optimizer was used.

5. Discussion

Landsliding is a very complicated process. So far, many scholars have

Fig. 12. The prediction rate curves by RNN and its variants. used different methods in LSM and compared their performance to
obtain the best model in the study area (Dou et al., 2019b; Thai Pham
et al., 2019). This study introduces a novel application of the RNN
framework for LSM. Under this framework, four advanced deep learning
Table 9
Chi-square statistics and significant levels between different methods. methods of RNN, LSTM, GRU and SRU were considered. Prior to land­
slide susceptibility analysis using a deep learning model, a primary step
Comparative pairs Chi-square (χ 2) p value Significance level
is to generate the sample data set containing positive and negative
RNN vs. LSTM 440.02 <0.001 Yes samples. Positive samples are prepared based on historical landslide
RNN vs. GRU 402.85 Yes
<0.001
inventory data, which are taken from the areas where landslides
RNN vs. SRU 229.31 <0.001 Yes
LSTM vs. GRU 444.49 <0.001 Yes occurred. There are several sampling methods for LSM, including seed
LSTM vs. SRU 298.49 <0.001 Yes cells (Süzen and Doyuran, 2004), single pixels (Atkinson and Massari,
GRU vs. SRU 300.18 <0.001 Yes 1998; Piacentini et al., 2012; Van Den Eeckhaut et al., 2006) and all
pixels (Ayalew and Yamagishi, 2005). In fact, the mixed type of land­
slides for deep learning modeling might cause misinterpretation and
occurred. In addition, both RNN and GRU successfully predicted this
error by using some sampling methods. In this study, however, the
area as a very high susceptible area, while LSTM and SRU predicted this
centroid point of the landslide polygon as positive data can significantly
area as a moderate susceptible area. For the sub-region (b), another
mitigate the negative effects of mixed landslide types. After constructing
landslide event in the satellite image of 2019, and RNN, LSTM and GRU
the sample data, choosing different landslide influencing factors plays
predicted this area as a highly susceptible area. For the sub-region (c), all
an important role in the final LSM results (Jaafari et al., 2019; Jiao et al.,
the RNN-related methods obtained satisfactory prediction results
2019). Therefore, it is necessary to evaluate the importance of these
because they predicted the landslide location as a highly susceptible
factors. First, the conditional independence between these factors in the
area. In summary, the proposed methods can accurately and effectively
study area was tested. Experimental results demonstrate that the abso­
predict the location of landslides, which is beneficial to disaster pre­
lute values of all the correlation coefficients between these factors are
vention and management.
below the critical threshold and the 16 selected factors are independent
of each other. Second, multi-collinearity analysis was employed to es­
4.4. Parameter analysis timate the correlation between these factors. Experimental results
proved that these factors have no multi-collinearity. Meanwhile, the AM
In our experiments, all the four models have built appropriate value obtained by the factor of lithology is higher than the other factors,
network structures and hidden units during the training process, so the indicating that this factor is the main cause of landslides in the study
trained models have better fitting capabilities. The model stabilizes the area, which is consistent with previous studies (Dou et al., 2019a; Hong
training by detecting the interaction of any possible situation. In the et al., 2018a).
process of network structures optimization, it is very important to find The regular RNN model is capable of making full use of contextual
suitable parameters and avoid the over-fitting problem. According to information to perform landslide spatial prediction. As for other modi­
our comprehensive experiments, the parameter settings of the four fied RNN models, the LSTM model implements the protection and
models have an impact on the network structure. The common param­ control of information through three structures: input gate, forgetting
eters of the four models are epoch, batch size, dropout, optimizer, loss gate and output gate, which interact in a very special way. The GRU
function and learn rate. In this section, the effects of several critical model is simpler than the LSTM model because it combines input and
parameters will be analyzed to better understand the application of forget gates into a single update gate. In addition, it has an additional
RNNs in LSM. reset gate. The SRU is set up with forgot and reset gates. The RNN-
Fig. 14 shows the AUC values by the RNNs under different parame­ related models were evaluated and analyzed using the verification set.
ters for the study area. The parameter of epoch refers to all samples Experimental results showed that GRU achieved the highest AUC and
involved in the training process in the training set. To obtain its optimal recall values. This is because GRU is a heuristic method that has proven
value, the number of epochs was fixed from 50 to 500 with a step size of its effectiveness in many fields (Chung et al., 2014a, 2015; Tang et al.,
50. In Fig. 14 (a), RNN achieved the highest AUC value with an epoch of 2015). Moreover, GRU can achieve better predictive capabilities with a
50. For the other three models, the AUC value first demonstrates an simple structure than the other models, which can reduce the compu­
upward trend. As the number of epochs is further increased, the AUC tation cost and improve its effectiveness. For these landslide suscepti­
values decrease. The parameter of batch size refers to the number of bility maps, it can be observed from Table 7 that all susceptibility maps
training samples calculated during the optimization process, and show similar LD distribution trends, which is consistent with previous

14
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Fig. 13. Prediction performance assessment using two high-resolution remote sensing images.

studies (Juliev et al., 2019; Ko and Lo, 2018; Yan et al., 2019). Moreover, in selecting parameters. By setting an appropriate learning rate, the
the effectiveness of the resultant susceptibility maps was further verified objective function can converge to a local minimum in an appropriate
using Gaogfen-1 satellite data. Specifically, comparing two different time. Therefore, setting the optimal parameters of the RNN model has a
temporal high-resolution images can show the validity of the resultant crucial impact on the optimization of the model and the experimental
susceptibility maps. Fig. 13 shows that landslide disasters did occur in results. Moreover, searching for optimized parameters is necessary and
these predicted high susceptible areas where no historical landslide important to perform a reliable landslide susceptibility map using the
existed, which indicates that our methods can accurately predicate the RNN techniques.
high risk areas of landslide disasters. In general, these deep learning
models are very effective for LSM, and the deep learning techniques 6. Conclusion
have broad prospects in landslide susceptibility assessment.
It was recorded that the parameters of the RNNs play a crucial role in Landslides are one of the most dangerous natural disasters in China,
demonstrating their predictive performance (Ghorbanzadeh et al., 2019; causing serious casualties and socioeconomic losses every year. This
Yu et al., 2017). Four key parameters were selected to discuss the impact paper mainly aims at the construction of LSM based on RNNs in Yongxin
on landslide susceptibility results, namely epoch, batch size, dropout County, China. A total of 364 historical landslide locations were
and optimizer. Within the range of values, each model has its optimal randomly divided into two parts: 70% for training and 30% for verifi­
parameters. The main function of setting the epoch is to divide the entire cation, and 16 influencing factors were chosen based on previous studies
training process of the model into several stages so that we can better and the techniques of the correlation matrix, multi-collinearity analysis,
observe and adjust the training of the model. The batch size can be IG and CF were used to evaluate the correlation between landslides and
regarded as a batch parameter, which controls the number of input these factors. The final landslide susceptibility maps were obtained by
samples that the deep learning network can process during the iteration. the four RNN-related models for comparison. From the experiment re­
Fig. 14 (b) shows that 32 and 64 are the optimal batch size. The results in sults, the two main conclusions can be summarized as follows. First, all
Fig. 14 (c) shows that the LSM results of GRU is better when 50% of the RNN-related models obtained satisfactory results in LSM. The AUC
neurons are eliminated during training. This is reasonable because values by these models were higher than 0.83 and the highest AUC value
setting the optimal dropout value can avoid the over-fitting problem was achieved by GRU. Second, the RNN-related models can produce
(Srivastava et al., 2014). Optimizer optimizes the weights of the con­ accurate landslide susceptibility maps in other prone zones with similar
structed neural network, and Adam is generally considered to be robust geo-environment. In the future, our research will focus on comparing the

15
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Fig. 14. Area under the curve (AUC) values by RNNs under different parameters for the study area. (a) Epoch, (b) batch size, (c) dropout and (d) optimizer.

RNN-based methods with other benchmark models and deep learning Acknowledgments
models in the field of LSM.
This work was supported by the National Natural Science Foundation
Data availability of China (61271408, 41602362). The authors acknowledge the joint
PhD scholarship awarded to Haoyuan Hong (201906860029) supported
All source codes related to this article can be found at https://github. by the China Scholarship Council. The authors would also like to thank
com/xmblb/RNN.git, an open-source online data repository hosted at the associate editor Candan Gokceoglu and the three anonymous re­
GitHub. viewers for their valuable comments and suggestions, which signifi­
cantly improved the quality of this paper.
Author Contributions
Appendix A. Supplementary data
Yi Wang had the original idea of this study, finished the first draft
and supervised the research, Zhice Fang and Mao Wang coded the Supplementary data to this article can be found online at https://doi.
research project and conducted the experiments, Ling Peng and org/10.1016/j.cageo.2020.104445.
Haoyuan Hong collected and analyzed research material and contrib­
uted to the editing and review of the manuscript. References

Declaration of competing interest Atkinson, P.M., Massari, R., 1998. Generalised linear modelling of susceptibility to
landsliding in the central Apennines, Italy. Comput. Geosci. 24, 373–385.
Ayalew, L., Yamagishi, H., 2005. The application of GIS-based logistic regression for
The authors declare that they have no known competing financial landslide susceptibility mapping in the Kakuda-Yahiko Mountains, Central Japan.
interests or personal relationships that could have appeared to influence Geomorphology 65, 15–31.
Bagui, S.C., 2005. Combining pattern classifiers: methods and algorithms. Technometrics
the work reported in this paper. 47, 517–518.
Basu, T., Pal, S., 2018. RS-GIS based morphometrical and geological multi-criteria
approach to the landslide susceptibility mapping in Gish River Basin, West Bengal,
India. Adv. Space Res. 63, 1253–1269.

16
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

Booth, G.D., Niccolucci, M.J., Schuster, E.G., 1994. Identifying Proxy Sets in Multiple Hong, H., Liu, J., Bui, D.T., Pradhan, B., Acharya, T.D., Pham, B.T., Zhu, A.-X., Chen, W.,
Linear Regression: an Aid to Better Coefficient Interpretation. Research paper INT Ahmad, B.B., 2018a. Landslide susceptibility mapping using J48 decision tree with
(USA). AdaBoost, bagging and rotation forest ensembles in the guangchang area (China).
Bui, D.T., Ho, T.-C., Pradhan, B., Pham, B.-T., Nhu, V.-H., Revhaug, I., 2016. GIS-based Catena 163, 399–413.
modeling of rainfall-induced landslides using data mining-based functional trees Hong, H., Pradhan, B., Sameen, M.I., Kalantar, B., Zhu, A., Chen, W., 2018b. Improving
classifier with AdaBoost, Bagging, and MultiBoost ensemble frameworks. Environ. the accuracy of landslide susceptibility model using a novel region-partitioning
Earth Sci. 75, 1101. approach. Landslides 15, 753–772.
Bui, D.T., Tuan, T.A., Hoang, N.-D., Thanh, N.Q., Nguyen, D.B., Van Liem, N., Jaafari, A., Panahi, M., Pham, B.T., Shahabi, H., Bui, D.T., Rezaie, F., Lee, S., 2019. Meta
Pradhan, B., 2017. Spatial prediction of rainfall-induced landslides for the Lao Cai optimization of an adaptive neuro-fuzzy inference system with grey wolf optimizer
area (Vietnam) using a hybrid intelligent approach of least squares support vector and biogeography-based optimization algorithms for spatial prediction of landslide
machines inference model and artificial bee colony optimization. Landslides 14, susceptibility. Catena 175, 430–445.
447–458. Jiang, C., Chen, S., Chen, Y., Bo, Y., Han, L., Guo, J., Feng, Z., Zhou, H., 2018.
Bui, Q.-T., Nguyen, Q.-H., Nguyen, X.L., Pham, V.D., Nguyen, H.D., Pham, V.-M., 2019. Performance analysis of a deep simple recurrent unit recurrent neural network (SRU-
Verification of novel integrations of swarm intelligence algorithms into deep RNN) in MEMS gyroscope de-noising. Sensors 18, 4471.
learning neural network for flood susceptibility mapping. J. Hydrol 124379. Jiao, Y., Zhao, D., Ding, Y., Liu, Y., Xu, Q., Qiu, Y., Liu, C., Liu, Z., Zha, Z., Li, R., 2019.
Chen, H., Zeng, Z., Tang, H., 2015. Landslide deformation prediction based on recurrent Performance evaluation for four GIS-based models purposed to predict and map
neural network. Neural Process. Lett. 41, 169–178. landslide susceptibility: a case study at a World Heritage site in Southwest China.
Chen, W., Zhang, S., Li, R., Shahabi, H., 2018. Performance evaluation of the GIS-based Catena 183, 104221.
data mining techniques of best-first decision tree, random forest, and naïve Bayes Juliev, M., Mergili, M., Mondal, I., Nurtaev, B., Pulatov, A., Hübl, J., 2019. Comparative
tree for landslide susceptibility modeling. Sci. Total Environ. 644, 1006–1018. analysis of statistical methods for landslide susceptibility mapping in the Bostanlik
Cho, K., Van Merri€enboer, B., Bahdanau, D., Bengio, Y., 2014. On the Properties of District, Uzbekistan. Sci. Total Environ. 653, 801–814.
Neural Machine Translation: Encoder-Decoder Approaches arXiv preprint arXiv: Ko, F.W., Lo, F.L., 2018. From landslide susceptibility to landslide frequency: a territory-
1409.1259. wide study in Hong Kong. Eng. Geol. 242, 12–22.
Chung, C.-J.F., Fabbri, A.G., 1993. The representation of geoscience information for data Kornejady, A., Bahremand, A., Ownegh, M.J.C., 2017a. Landslide susceptibility
integration. Nonrenewable Resour. 2, 122–139. assessment using maximum entropy model with two different data sampling
Chung, J., Gulcehre, C., Cho, K., Bengio, Y., 2014a. Empirical Evaluation of Gated methods. Catena 152, 144–162.
Recurrent Neural Networks on Sequence Modeling (arXiv: Neural and Evolutionary Kornejady, A., Ownegh, M., Rahmati, O., Bahremand, A., 2017b. Landslide susceptibility
Computing). assessment using three bivariate models considering the new topo-hydrological
Chung, J., Gulcehre, C., Cho, K., Bengio, Y., 2014b. Empirical Evaluation of Gated factor: hand. Geocarto Int. 33, 1–19.
Recurrent Neural Networks on Sequence Modeling arXiv preprint arXiv:1412.3555. Lee, S., Pradhan, B., 2007. Landslide hazard mapping at Selangor, Malaysia using
Chung, J., Gulcehre, C., Cho, K., Bengio, Y., 2015. Gated Feedback Recurrent Neural frequency ratio and logistic regression models. Landslides 4, 33–41.
Networks (arXiv: Neural and Evolutionary Computing). Lei, T., Zhang, Y., Wang, S.I., Dai, H., Artzi, Y., 2018. Simple recurrent units for highly
Clerici, A., Perego, S., Tellini, C., Vescovi, P., 2006. A GIS-based automated procedure for parallelizable recurrence. In: Proceedings of the 2018 Conference on Empirical
landslide susceptibility mapping by the conditional analysis method: the Baganza Methods in Natural Language Processing, pp. 4470–4481.
valley case study (Italian Northern Apennines). Environ. Geol. 50, 941–961. Matthews, B.W., 1975. Comparison of the predicted and observed secondary structure of
Crozier, M.J., 2018. Reprint of “A proposed cell model for multiple-occurrence regional T4 phage lysozyme. Biochim. Biophys. Acta Protein Struct. 405, 442–451.
landslide events: implications for landslide susceptibility mapping”. Geomorphology Meusburger, K., Alewell, C., 2009. On the influence of temporal change on the validity of
307, 3–11. landslide susceptibility maps. Natural Hazards and Earth System Sciences (NHESS) &
Ding, Q., Chen, W., Hong, H., 2017. Application of frequency ratio, weights of evidence Discussions (NHESSD) 9, 1495–1507.
and evidential belief function models in landslide susceptibility mapping. Geocarto Mutlu, B., Nefeslioglu, H.A., Sezer, E.A., Akcayol, M.A., Gokceoglu, C., 2019. An
Int. 32, 619–639. experimental research on the use of recurrent neural networks in landslide
Donahue, J., Anne Hendricks, L., Guadarrama, S., Rohrbach, M., Venugopalan, S., susceptibility mapping. ISPRS Int. J. Geo-Inf. 8, 578.
Saenko, K., Darrell, T., 2015. Long-term recurrent convolutional networks for visual Nepal, N., Chen, J., Chen, H., Sharma, T.P.P., 2019. Assessment of landslide
recognition and description. Proc. IEEE.Conf. Comput. Vis. Pattern Recogn. susceptibility along the araniko highway in poiqu/bhote koshi/sun koshi watershed,
2625–2634. Nepal himalaya. Progress in Disaster Science 3, 100037.
Dou, J., Yunus, A.P., Bui, D.T., Merghadi, A., Sahana, M., Zhu, Z., Chen, C.-W., Neuh€ auser, B., Geomorphology, B.T.J., 2007. Landslide susceptibility assessment using
Khosravi, K., Yang, Y., Pham, B.T., 2019a. Assessment of advanced random forest “weights-of-evidence” applied to a study area at the Jurassic escarpment (SW-
and decision tree algorithms for modeling rainfall-induced landslide susceptibility in Germany). Geomorphology 86, 12–24.
the Izu-Oshima Volcanic Island, Japan. Sci. Total Environ. 662, 332–346. Nicu, I.C., As�andulesei, A., 2018. GIS-based evaluation of diagnostic areas in landslide
Dou, J., Yunus, A.P., Tien Bui, D., Sahana, M., Chen, C.-W., Zhu, Z., Wang, W., Pham, B. susceptibility analysis of Bahluieț River Basin (Moldavian Plateau, NE Romania). Are
T., 2019b. Evaluating GIS-based multiple statistical models and data mining for Neolithic sites in danger? Geomorphology 314, 27–41.
earthquake and rainfall-induced landslide susceptibility using the LiDAR DEM. Rem. Noda, K., Yamaguchi, Y., Nakadai, K., Okuno, H.G., Ogata, T., 2015. Audio-visual speech
Sens. 11, 638. recognition using deep learning. Appl. Intell. 42, 722–737.
Dou, J., Yunus, A.P., Xu, Y., Zhu, Z., Chen, C.-W., Sahana, M., Khosravi, K., Yang, Y., Nohani, E., Moharrami, M., Sharafi, S., Khosravi, K., Pradhan, B., Pham, B.T., Lee, S., M
Pham, B.T., 2019c. Torrential rainfall-triggered shallow landslide characteristics and Melesse, A., 2019. Landslide susceptibility mapping using different GIS-based
susceptibility assessment using ensemble data-driven models in the Dongjiang bivariate models. Water 11, 1402.
Reservoir Watershed, China. Nat. Hazards 97, 579–609. Nourani, V., Pradhan, B., Ghaffari, H., Sharifi, S.S.J.N.H., 2014. Landslide susceptibility
Ercanoglu, M., Gokceoglu, C., 2002. Assessment of landslide susceptibility for a mapping at Zonouz Plain, Iran using genetic programming and comparison with
landslide-prone area (north of Yenice, NW Turkey) by fuzzy approach. Environ. frequency ratio, logistic regression, and artificial neural network models. Nat.
Geol. 41, 720–730. Hazards 71, 523–547.
Erener, A., Sivas, A.A., Selcuk-Kestel, A.S., Düzgün, H.S., 2017. Analysis of training Nsengiyumva, J.B., Luo, G., Hakorimana, E., Mind’je, R., Gasirabo, A., Mukanyandwi, V.,
sample selection strategies for regression-based quantitative landslide susceptibility 2019. Comparative analysis of deterministic and semiquantitative approaches for
mapping methods. Comput. Geosci. 104, 62–74. shallow landslide risk modeling in Rwanda. Risk Anal. 39, 2576–2595.
Galli, M., Ardizzone, F., Cardinali, M., Guzzetti, F., Reichenbach, P., 2008. Comparing O’brien, R.M., 2007. A caution regarding rules of thumb for variance inflation factors.
landslide inventory maps. Geomorphology 94, 268–289. Qual. Quantity 41, 673–690.
Gao, H., 2011. Analyses on coupling correlation between landslides and rainfall and Oh, H.-J., Pradhan, B., 2011. Application of a neuro-fuzzy model to landslide-
application in the early-warning and forecast. In: 2011 International Conference on susceptibility mapping for shallow landslides in a tropical hilly area. Comput.
Electric Technology and Civil Engineering (ICETCE), pp. 6432–6436. Geosci. 37, 1264–1276.
Ghorbanzadeh, O., Blaschke, T., Gholamnia, K., Meena, S.R., Tiede, D., Aryal, J., 2019. Ozer, B., Mutlu, B., Nefeslioglu, H., Sezer, E., Rouai, M., Dekayir, A., Gokceoglu, C.,
Evaluation of different machine learning methods and deep-learning convolutional 2019. On the use of hierarchical fuzzy inference systems (HFIS) in expert-based
neural networks for landslide detection. Rem. Sens. 11, 196. landslide susceptibility mapping: the central part of the Rif Mountains (Morocco).
G€
okceoglu, C., Aksoy, H., 1996. Landslide susceptibility mapping of the slopes in the Bull. Eng. Geol. Environ. 1–18.
residual soils of the Mengen region (Turkey) by deterministic stability analyses and Pan, Y., Mei, T., Yao, T., Li, H., Rui, Y., 2016. Jointly modeling embedding and
image processing techniques. Eng. Geol. 44, 147–161. translation to bridge video and language. Proc. IEEE.Conf. Comput. Vis. Pattern
Graves, A., 2012. Long short-term memory. In: Graves, A. (Ed.), Supervised Sequence Recogn. 4594–4602.
Labelling with Recurrent Neural Networks. Springer Berlin Heidelberg, Berlin, Parry, S., 2011. Chapter fifteen - the application of geomorphological mapping in the
Heidelberg, pp. 37–45. assessment of landslide hazard in Hong Kong. In: Smith, M.J., Paron, P., Griffiths, J.
Guzzetti, F., Mondini, A.C., Cardinali, M., Fiorucci, F., Santangelo, M., Chang, K.-T., S. (Eds.), Developments in Earth Surface Processes. Elsevier, pp. 413–441.
2012. Landslide inventory maps: new tools for an old problem. Earth Sci. Rev. 112, Peshevski, I., Jovanovski, M., Abolmasov, B., Papi�c, J., Đuri�c, U., Marjanovi�c, M.,
42–66. Haque, U., Nedelkovska, N., 2019. Preliminary regional landslide susceptibility
Hinton, G.E., Salakhutdinov, R.R., 2006. Reducing the dimensionality of data with neural assessment using limited data. Geol. Croat. 72, 81–92.
networks. Science 313, 504–507. Pham, B.T., 2018. A novel classifier based on composite hyper-cubes on iterated random
Hoermann, S., Bach, M., Dietmayer, K., 2018. Dynamic occupancy grid prediction for projections for assessment of landslide susceptibility. J. Geol. Soc. India 91,
urban autonomous driving: a deep learning approach with fully automatic labeling. 355–362.
In: 2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, Pham, B.T., Bui, D.T., Dholakia, M., Prakash, I., Pham, H.V., 2016. A comparative study
pp. 2056–2063. of least square support vector machines and multiclass alternating decision trees for

17
Y. Wang et al. Computers and Geosciences 138 (2020) 104445

spatial prediction of rainfall-induced landslides in a tropical cyclones area. Geotech. trees classifier with AdaBoost, Bagging, and MultiBoost ensemble frameworks.
Geol. Eng. 34, 1807–1824. Environ. Earth Sci. 75, 1101.
Pham, B.T., Bui, D.T., Dholakia, M.B., Prakash, I., Pham, H.V., Mehmood, K., Le, H.Q., Tsangaratos, P., Ilia, I., Hong, H., Chen, W., Xu, C., 2017. Applying Information Theory
2017a. A novel ensemble classifier of rotation forest and Naïve Bayer for landslide and GIS-based quantitative methods to produce landslide susceptibility maps in
susceptibility assessment at the Luc Yen district, Yen Bai Province (Viet Nam) using Nancheng County, China. Landslides 14, 1091–1111.
GIS. Geomatics, Nat. Hazards Risk 8, 649–671. Van Den Eeckhaut, M., Vanwalleghem, T., Poesen, J., Govers, G., Verstraeten, G.,
Pham, B.T., Bui, D.T., Prakash, I., Dholakia, M.B.J.C., 2017b. Hybrid integration of Vandekerckhove, L., 2006. Prediction of landslide susceptibility using rare events
Multilayer Perceptron Neural Networks and machine learning ensembles for logistic regression: a case-study in the Flemish Ardennes (Belgium). Geomorphology
landslide susceptibility assessment at Himalayan area (India) using GIS. Catena 149, 76, 392–410.
52–63. Wang, L.-J., Guo, M., Sawada, K., Lin, J., Zhang, J., 2016. A comparative study of
Pham, B.T., Prakash, I., Dou, J., Singh, S.K., Trinh, P.T., Tran, H.T., Le, T.M., Van landslide susceptibility maps using logistic regression, frequency ratio, decision tree,
Phong, T., Khoi, D.K., Shirzadi, A., 2019a. A novel hybrid approach of landslide weights of evidence and artificial neural network. Geosci. J. 20, 117–136.
susceptibility modelling using rotation forest ensemble and different base classifiers. Wang, Q., Wang, Y., Niu, R., Peng, L., 2017. Integration of information theory, K-means
Geocarto Int. 1–25. cluster analysis and the logistic regression model for landslide susceptibility
Pham, B.T., Prakash, I., Khosravi, K., Chapi, K., Trinh, P.T., Ngo, T.Q., Hosseini, S.V., mapping in the Three Gorges Area, China. Rem. Sens. 9, 938.
Bui, D.T., 2019b. A comparison of Support Vector Machines and Bayesian algorithms Wang, Y., Duan, H., Hong, H., 2019a. A comparative study of composite kernels for
for landslide susceptibility modelling. Geocarto Int. 34, 1385–1407. landslide susceptibility mapping: a case study in Yongxin County, China. Catena 183,
Piacentini, D., Troiani, F., Soldati, M., Notarnicola, C., Savelli, D., Schneiderbauer, S., 104217.
Strada, C., 2012. Statistical analysis for assessing shallow-landslide susceptibility in Wang, Y., Fang, Z., Hong, H., 2019b. Comparison of convolutional neural networks for
South Tyrol (south-eastern Alps, Italy). Geomorphology 151, 196–206. landslide susceptibility mapping in Yanshan County, China. Sci. Total Environ. 666,
Pourghasemi, H., Pradhan, B., Gokceoglu, C., Moezzi, K.D., 2012a. Landslide 975–993.
Susceptibility Mapping Using a Spatial Multi Criteria Evaluation Model at Haraz Wang, Y., Fang, Z., Hong, H., Peng, L., 2019c. Flood susceptibility mapping using
Watershed, Iran, Terrigenous Mass Movements. Springer, pp. 23–49. convolutional neural network frameworks. J. Hydrol 582, 124482.
Pourghasemi, H.R., Pradhan, B., Gokceoglu, C., 2012b. Application of fuzzy logic and Xiao, L., Zhang, Y., Peng, G., 2018. Landslide susceptibility assessment using integrated
analytical hierarchy process (AHP) to landslide susceptibility mapping at Haraz deep learning algorithm along the China-Nepal Highway. Sensors 18, 4436.
watershed, Iran. Nat. Hazards 63, 965–996. Xu, S., Niu, R., 2018. Displacement prediction of Baijiabao landslide based on empirical
Powers, D.M., 2011. Evaluation: from Precision, Recall and F-Measure to ROC, mode decomposition and long short-term memory neural network in Three Gorges
Informedness, Markedness and Correlation. area, China. Comput. Geosci. 111, 87–96.
Regmi, A.D., Yoshida, K., Pourghasemi, H.R., DhitaL, M.R., Pradhan, B., 2014. Landslide Yalcin, A., 2008. GIS-based landslide susceptibility mapping using analytical hierarchy
susceptibility mapping along Bhalubang—shiwapur area of mid-Western Nepal using process and bivariate statistics in Ardesen (Turkey): comparisons of results and
frequency ratio and conditional probability models. J. Mt. Sci. 11, 1266–1285. confirmations. Catena 72, 1–12.
Regmi, N.R., Giardino, J.R., Geomorphology, J.D.V.J., 2010. Modeling susceptibility to Yan, F., Zhang, Q., Ye, S., Ren, B., 2019. A novel hybrid approach for landslide
landslides using the weight of evidence approach: western Colorado, USA. susceptibility mapping integrating analytical hierarchy process and normalized
Geomorphology 115, 172–187. frequency ratio methods with the cloud model. Geomorphology 327, 170–187.
Sevgen, E., Kocaman, S., Nefeslioglu, H.A., Gokceoglu, C., 2019. A novel performance Yang, J., Song, C., Yang, Y., Xu, C., Guo, F., Xie, L., 2019. New method for landslide
assessment approach using photogrammetric techniques for landslide susceptibility susceptibility mapping supported by spatial logistic regression and GeoDetector: a
mapping with logistic regression, ANN and random forest. Sensors 19, 3940. case study of Duwen Highway Basin, Sichuan Province, China. Geomorphology 324,
Shirzadi, A., Soliamani, K., Habibnejhad, M., Kavian, A., Chapi, K., Shahabi, H., 62–71.
Chen, W., Khosravi, K., Thai Pham, B., Pradhan, B., 2018b. Novel GIS based machine Yu, S., Jia, S., Xu, C., 2017. Convolutional neural networks for hyperspectral image
learning algorithms for shallow landslide susceptibility mapping. Sensors 18, 3777. classification. Neurocomputing 219, 88–98.
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R., 2014. Yu, X., Wang, Y., Niu, R., Hu, Y., 2016. A combination of geographically weighted
Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. regression, particle swarm optimization and support vector machine for landslide
Res. 15, 1929–1958. susceptibility mapping: a case study at wanzhou in the three gorges area, China. Int.
Steger, S., Brenning, A., Bell, R., Petschko, H., Glade, T., 2016. Exploring discrepancies J. Environ. Res. Publ. Health 13, 487.
between quantitative validation results and the geomorphic plausibility of statistical Z^ezere, J., Pereira, S., Melo, R., Oliveira, S., Garcia, R., 2017. Mapping landslide
landslide susceptibility maps. Geomorphology 262, 8–23. susceptibility using data-driven methods. Sci. Total Environ. 589, 250–267.
Süzen, M.L., Doyuran, V., 2004. Data driven bivariate landslide susceptibility assessment Zhang, G., Wang, M., Liu, K., 2019. Forest fire susceptibility modeling using a
using geographical information systems: a method and application to Asarsuyu convolutional neural network for yunnan Province of China. Int. J.Disaster Risk.Sci.
catchment. Turkey. Eng. Geol. 71, 303–321. 10, 386–403.
Tallarida, R.J., Murray, R.B., 1987. Chi-square test. In: Tallarida, R.J., Murray, R.B. Zhou, C., Yin, K., Cao, Y., Ahmed, B., Li, Y., Catani, F., Pourghasemi, H.R., 2018.
(Eds.), Manual of Pharmacologic Calculations: with Computer Programs. Springer Landslide susceptibility modeling applying machine learning methods: a case study
New York, New York, NY, pp. 140–142. from Longju in the Three Gorges Reservoir area, China. Comput. Geosci. 112, 23–37.
Tang, D., Qin, B., Liu, T., 2015. Document modeling with gated recurrent neural network Zhu, A.-X., Miao, Y., Liu, J., Bai, S., Zeng, C., Ma, T., Hong, H., 2019. A similarity-based
for sentiment classification. In: Proceedings of the 2015 Conference on Empirical approach to sampling absence data for landslide susceptibility mapping using data-
Methods in Natural Language Processing, pp. 1422–1432. driven methods. Catena 183, 104188.
Thai Pham, B., Bui, D.T., Prakash, I., 2019. Landslide susceptibility modelling using Zhu, A.-X., Miao, Y., Yang, L., Bai, S., Liu, J., Hong, H., 2018. Comparison of the
different advanced decision trees methods. Civ. Eng. Environ. Syst. 35, 139–157. presence-only method and presence-absence method in landslide susceptibility
Tien Bui, D., Ho, T.-C., Pradhan, B., Pham, B.-T., Nhu, V.-H., Revhaug, I., 2016. GIS- mapping. Catena 171, 222–233.
based modeling of rainfall-induced landslides using data mining-based functional

18