Access to this full-text is provided by De Gruyter.
Content available from Open Geosciences
This content is subject to copyright. Terms and conditions apply.
Research Article
Fei Teng, Yimin Mao, Yican Li*, Subin Qian, and Yaser A. Nanehkaran*
Comparative models of support-vector machine,
multilayer perceptron, and decision tree
predication approaches for landslide
susceptibility analysis
https://doi.org/10.1515/geo-2022-0642
received February 03, 2024; accepted April 13, 2024
Abstract: Naqadeh Region (NR) is one of the most sensitive
regions regarding geo-hazards occurrence in Northwest of
Iran. The landslides triggering parameters that identified
for the studied region are classified as elevation, aspect,
slope angle, lithology, drainage density, distance to river,
weathering, land-cover, precipitation, vegetation, distance
to faults, distance to roads, and distance to the cities. These
triggering factors are selected based on conducting field
survey, remote-sensing investigation, and historical devel-
opment background assessment. Regarding the investiga-
tions, 12 large-scale, 15 medium-scale, and 30 small-scale
historical landslides (57 in total) were recorded in the
NR. The historical landslides were used to provide sensitive
area with high probability of ground movements. The
objectives of this study are multifaceted, aiming to address
critical gaps in understanding and predicting landslide
susceptibility in the NR. First, the study seeks to evaluate
and compare the effectiveness of support-vector machine
(SVM), multilayer perceptron (MLP), and decision tree (DT)
algorithms in predicting landslide susceptibility. So, as meth-
odology, the presented study used comparative models
for landslide susceptibility based on SVM, MLP, and DT
approaches. The predictive models were compared based
on model accuracy as the area under the curve of the receiver
operating characteristic curve. According to the estimated
results, MLP is the highest rank of overall accuracy to provide
susceptibility maps for landslides in NR. From a perspective
of the risk ability, the west and south-west sides of the county
were identified within the hazard area.
Keywords: landslide susceptibility, support-vector machine,
multilayer perceptron, decision tree, geo-hazards
1 Introduction
Landslides is one of the important geo-hazards were
responsible various scale ground deformation, financial
damages, and live losses [1] which are the ground move-
ment, and debris down a slope or low topographic levels
[2]. They can be caused by natural events such as heavy
rainfall, earthquakes, volcanic activity, erosion, or human
activities such as construction or mining [3]. Landslides
can range in size and severity from small, localized events
to large-scale disasters that can cause significant damage to
property and infrastructure [4]. There are several different
types of landslides based on Varnes [5] and Highland and
Bobrowsky [6] including rockfalls, debris flows, toppling, and
sliding. Regardless of the type of landslides can have significant
impacts on the environment, eco-systems, and human lives [7].
They can also have economic and social impacts, particularly
in areas where infrastructure or communities are in landslide-
prone areas. Comprehending the origins and triggering factors
behind landslides occurrence, alongside implementing efficient
management tactics to minimize their repercussions, constitutes
asignificant focal point for geologists and geoengineers [8].
Landslide susceptibility refers to the potential of an
area or region to experience landslides which is deter-
mined by a combination of various triggering factors
which involves identifying areas that are likely to experi-
ence landslides in the future. This information is essential
for land-use planning, hazard mitigation, and risk reduc-
tion [9,10]. Several methods are used to assess landslide
Fei Teng, Yimin Mao: School of Information and Engineering, Shaoguan
University, Shaoguan 512005, Guangdong, China
* Corresponding author: Yican Li, Xi’an Center of Mineral Resources
Survey, China Geological Survey, Xi’an, 710105, Shaanxi, China,
e-mail: canmermer@126.com
Subin Qian: School of Information Engineering, Yancheng Teachers
University, Yancheng 224002, Jiangsu, China
* Corresponding author: Yaser A. Nanehkaran, School of Information
Engineering, Yancheng Teachers University, Yancheng 224002, Jiangsu,
China, e-mail: yaser@yctu.edu.cn
Open Geosciences 2024; 16: 20220642
Open Access. © 2024 the author(s), published by De Gruyter. This work is licensed under the Creative Commons Attribution 4.0 International License.
susceptibility which is classified into the quantitative,
qualitative, and semi-quantitative procedures [11]. These
general groups can be divided into the various sub-groups
containing various methods such as deterministic, statistic,
probabilistic, heuristic, geostatistic, inventory, and knowledge-
based approaches. Each approach has its own advantages and
disadvantages to provide the landslide susceptibility maps
[12,13]. According to Ercanoglu and Gokceoglu [14], landslide
susceptibility assessment has been on the rise worldwide,
but there is currently no established overall standard pro-
cedure for doing so. Thus, each researcher follows their
specific methods to achieve more accurate susceptibility
analysis for landslides.
With recent developments and advancements in com-
puter-based modeling, the landslide susceptibility assess-
ments become more straightforward and provide highly
accurate/reliable results [3]. In the meantime, the artificial
intelligence (AI)-based method received high rate of suc-
cess in the susceptibility analysis of landslides [12]. There
are several advantages of AI-based methods for landslide
susceptibility assessment which can be stated as accurate
predictions, faster processing, cost-effective, adaptability,
improved decision-making, reduced risk, multidisciplinary
integration, objective analysis, scalability, automation, and
accessibility, constantly evolving and improving efficiently
[15,16]. While there are many advantages to using AI-based
landslide susceptibility assessment, there are also some
limitations that need to be considered [10].
AI-based methods offer several advantages for employing
support-vector machine (SVM), multilayer perceptron (MLP),
and decision tree (DT) algorithms in predicting landslide
susceptibility in the Naqadeh Region (NR). First, these AI algo-
rithms can effectively handle complex and nonlinear relation-
ships within the dataset, which is crucial for capturing the
intricate interplay of various factors influencing landslide
occurrence [11]. SVMs excel in identifying optimal decision
boundaries, making them well-suited for distinguishing
between different classes of landslide susceptibility. MLPs,
with their ability to learn hierarchical representations of
data, can capture subtle patterns and dependencies, enhan-
cing the accuracy of landslide susceptibility models [12]. DT
algorithms provide interpretable decision rules, enabling
stakeholders to understand the factors driving landslide sus-
ceptibility in the NR region and facilitating informed deci-
sion-making for risk mitigation strategies. Moreover, AI
algorithms offer scalability and adaptability, allowing for
the integration of diverse datasets and variables relevant
to landslide susceptibility assessment in the NR region.
These algorithms can efficiently process large volumes of
spatial and environmental data, including elevation, slope
angle, lithology, precipitation, and land cover, enabling a
comprehensive analysis of landslide triggers [13]. By lever-
aging AI techniques, researchers can develop more robust
and reliable landslide susceptibility models tailored to the
specific characteristics of the NR region, thereby enhancing the
effectiveness of landslide risk management efforts. Further-
more, AI algorithms facilitate continuous learning and improve-
ment of landslide susceptibility models over time [14,15].
Through iterative refinement and optimization, SVM, MLP,
and DT algorithms can adapt to evolving environmental
conditions and incorporate new data sources, enhancing
the accuracy and reliability of landslide susceptibility pre-
dictions for the NR region. This dynamic capability ensures
that landslide risk assessment remains up-to-date and respon-
sive to changing environmental factors, enabling stakeholders
to proactively mitigate landslide hazards and safeguard com-
munities and infrastructure in the NR.
The objectives of this study are multifaceted, aiming to
address critical gaps in understanding and predicting land-
slide susceptibility in the NR. By assessing the performance
of these AI techniques, the study aims to identify the most
suitable modeling approach for accurately characterizing
landslide-prone areas in the NR. Additionally, the study
aims to enhance the interpretability of landslide suscept-
ibility models by analyzing the contributing factors and
spatial patterns identified by each algorithm, thereby pro-
viding valuable insights for stakeholders and decision-
makers involved in landslide risk management. So, the
presented study used a comparative evaluation based on
well-known machine learning classifiers including SVM,
MLP, and DT. These classifiers have good capabilities in
landslide susceptibility assessments, which are used for
various cases worldwide. According to the various contri-
butions that published in different literatures, it can be
stated that the main advantages of the SVM, MLP, and DT
in landslide susceptibility analysis are categorized as fol-
lows [17–20].
SVM is effective in high-dimensional spaces and is
good at separating data that are not linearly separable.
This classifier has a regularization parameter that helps
prevent overfitting, making it less prone to errors caused
by noise or outliers in the data. Also, SVM can handle both
linear and non-linear classification and regression tasks.
MLP is a powerful model for non-linear classification and
regression tasks. This classifier can learn complex decision
boundaries by using multiple hidden layers of neurons and
handle both continuous and categorical data, making it a
versatile algorithm. Also, MLP is relatively easy to use and
implement, and it can be trained using a variety of optimi-
zation techniques. DT is easy to understand and interpret,
making it useful for explaining decisions to non-technical
stakeholders. This classifier can handle both numerical
2Fei Teng et al.
and categorical data, making it a versatile algorithm and
capture non-linear relationships between variables.
Lee et al. [21] employed weighted artificial neural net-
works (ANNs) to assess landslide susceptibility, offering a
valuable approach for spatial data handling and proces-
sing. They utilized a probabilistic method to determine
the learning rate for factors contributing to landslide
occurrence. These findings were then leveraged to con-
struct an ANN-driven landslide susceptibility index. In a
similar vein, Ermini et al. [22] utilized MLP and the prob-
abilistic neural network, both belonging to the shallow
learning category, to evaluate landslide susceptibility in
the Riomaggiore catchment, a subwatershed situated within
the Northern Apennines’Reno River basin in Italy. Kanungo
et al. [23], Oh and Pardhan [24], Quan and Lee [25], Park et al.
[26], and Nourani et al. [27] conducted comparative analyses
aimed at forecasting the likelihood of landslide occurrences
in mountainous regions, particularly in tropical areas. Their
approach involved integrating benchmark classifiers and
ANN techniques. The common thread among these studies
was the emphasis on enhancing the main database for
training and testing sets, along with diversifying benchmark
learning methodologies, such as frequency ratio, logistic
regression, adaptive neuro-fuzzy inference system, and ana-
lytic hierarchy process. In contrast, Liu and Wu [28], and Xiao
et al. [29] explored more sophisticated learning approaches for
landslide susceptibility mapping, yielding notable advance-
ments over traditional benchmark classifiers (referred to as
conventional machine learningorCML)andshallowlearning
methods. Furthermore, Ortiz and Martínez-Graña [30] and
Ghorbanzadeh et al. [31] leveraged convolutional neural
networks (CNNs) for landslide susceptibility assessments,
resulting in the creation of accurate susceptibility maps.
These researchers highlighted the favorable impact of
CNNs on satellite imagery, with outputs seamlessly transfer-
able into the geographic information system (GIS) environ-
ment. Mutlu et al. [32] employed recurrent neural networks
(RNNs) for susceptibility assessments and forecasting land-
slide-prone regions. While RNNs offer distinct advantages
over CNNs, particularly in accuracy and performance,
CNNs exhibit better adaptability for the task at hand.
The novelty of the presented study lies in its utilization
of comparative modeling techniques for landslide susceptibility
assessment. By employing SVM, MLP, and DT approaches,
the study provides a comprehensive evaluation of different
machine learning algorithms in predicting landslide sus-
ceptibility. This comparative approach allows for a thorough
analysis of the strengths and weaknesses of each model,
enabling researchers to identify the most suitable method
for landslide hazard assessment in the specific study area.
Furthermore, by assessing model accuracy using the area
under the curve (AUC) of the receiver operating character-
istic (ROC) curve, the study ensures a robust evaluation
criterion. The AUC-ROC metric provides a quantitative mea-
sure of model performance, considering both sensitivity and
specificity, thereby offering a more comprehensive assess-
ment of predictive capabilities. This approach enhances the
reliability and validity of the findings, facilitating informed
decision-making for landslide risk management and mitiga-
tion strategies in the NR and similar geographical contexts.
2 Study area
The NR is a district in Northwestern Iran, located in the
West Azerbaijan Province. The county encompasses the
city of Naqadeh and several surrounding towns and vil-
lages [33]. NR is situated on the bank of the Bayzawa River
and the Kani Sivin River, encompassing an old artificial
mound which is located is to the south-west of Urmia
Lake on the lower course of the Gadar river. It is situated
in the Zagros Mountains region, which is a complex moun-
tain range that spans western Iran. The region is charac-
terized by its rugged terrain, with steep slopes and deep
valleys, providing potential geo-hazards regarding land-
slides and rockfalls [34]. The location of the NR is illustrated
in Figure 1. In terms of climate, NR has a semi-arid to Med-
iterranean climate, with hot, dry summers and cool, wet
winters. The region receives most of its precipitation during
the winter months, with the summer months being rela-
tively dry [33]. As a general overview, the geography and
geology of the NR make it unique and complex regarding
geo-hazards and land-sliding, with rugged mountains, deep
valleys, and scenic rivers [34].
Also, by using a desk study check, 12 large-scale, 15
medium-scale, and 30 small-scale historical landslides (57
in total) were considered sensitive target points in the NR
to increase the accuracy of landslide susceptibility assess-
ment. The landslides chosen for analysis predominantly
encompass significant failures within NR. Additionally,
smaller-scale landslides frequently occur in proximity to
these larger events. By focusing on the primary large-scale
landslides, we aim to identify the most sensitive areas of
interest, which are pivotal in understanding potential cat-
astrophic impacts. So, the primary aim in selecting these
landslides is rooted in the historical geo disaster manage-
ment plan of the county, which underscores their notable
impact on the studied area. By identifying these factors and
their relative importance in an NR, it is possible to assess
the susceptibility of the region to landslides and to imple-
ment measures to mitigate the geo-hazard risk. These
Landslide susceptibility based on machine learning 3
measures can include slope stabilization techniques, drai-
nage systems, and land-use planning and management [11].
3 Geological setting
Geologically, the NR is affected by the Zagros fold-thrust
belt, which is a result of the collision between the Arabian
Plate and the Eurasian Plate. This collision has led to the
deformation and folding of the Earth’s crust, resulting in the
formation of the Zagros Mountains. The region is character-
ized by a variety of rock types, including limestone, sand-
stone, shale, and volcanic rocks. A geological map of the
studied region is provided in Figure 2. The rugged terrain
and geological features of the area make it susceptible to
landslides, particularly during periods of heavy rainfall or
seismic activity [35]. In recent years, several landslides have
occurred in NR, causing damage to infrastructures and
buildings. One notable landslide occurred in 2017 in the
village of Ahmadkhel, which is located near Naqadeh city
which is recorded as historical landslides. The landslide was
triggered by heavy rainfall, and it resulted in the death of six
people and the destruction of several homes [36]. To mitigate
landslides hazard effects, conducting a landslide suscept-
ibility assessment for the region and one of the key require-
ments to evaluate the level of risk, ongoing monitoring, and
mitigation efforts to minimize their impacts on communities
and the environment. This aim is targeted in this article.
4 Triggering factors
There are several techniques that are used to select and
identify the triggering factors of landslides, which are clas-
sified as historical development background check, remote-
sensing observations, field survey, and desk study check
[4,6]. To proceed with the extraction and selection of trig-
gering factors, the present study utilized a field survey, a
background check on historical development, and remote-
sensing observation. The landslide-triggering parameters
that are identified for the NR are classified as elevation,
aspect, slope angle, lithology, drainage density, distance to
river, weathering, land-cover, precipitation, vegetation, dis-
tance to faults, distance to roads, and distance to the cities
[6]. To elaborate further, the assessment process involved a
detailed examination of the past occurrences of landslides in
the area and their potential causes. Site investigations were
conducted to gather information on the geological charac-
teristics of the NR, as well as to identify the existing vegeta-
tion cover and its condition. Moreover, satellite imagery was
used to map the terrain and to assess the suitability of the
region for potential landslides. This analysis involved iden-
tifying the slope steepness and gradient, the presence of
potential unstable soil or rock formations, and the density
and condition of vegetation cover. By combining these eva-
luation methods, a thorough and comprehensive assessment
of each triggering factor was conducted to accurately eval-
uate the susceptibility of the NR to land-sliding probability.
Figure 1: Location of the studied region in Iran.
4Fei Teng et al.
Figure 3 depicts the rasterized triggering factors within
a GIS environment, which will be utilized to generate a
susceptibility map for the NR. The susceptibility map will
facilitate an accurate assessment of the likelihood of land-
slides in the NR, thereby enabling appropriate infrastruc-
tural and urban development planning. These triggering
factorial maps that prepared used as basic maps and infor-
mation layers for comparative modeling by SVM, MLP, and
DT. The gathered data will be used as the main database for
all prediction processes. In landslide susceptibility analysis
for NR, triggering parameters are key factors that contribute
to the initiation or occurrence of landslides. These parameters
interact with the inherent susceptibility of the terrain and
environmental conditions to increase the likelihood of slope
failure. Understanding and analyzing these triggering factors
are crucial for predicting and mitigating landslide hazards.
Let us explore how each of the mentioned triggering para-
meters relates to landslide occurrence in NR:
4.1 Precipitation
One of the most significant triggering factors is precipita-
tion. Heavy rainfall can saturate the soil, increase pore
water pressure, and reduce soil cohesion, making slopes
more prone to failure. Intense or prolonged rainfall events
can trigger landslides, especially in areas with steep slopes
and poorly drained soils.
4.2 Slope angle and aspect
Slope angle and aspect influence the stability of slopes.
Steeper slopes are inherently more susceptible to land-
slides, especially when combined with other factors such
as weak lithology or intense weathering. Aspect can affect
the exposure of slopes to sunlight and rainfall, influencing
rates of weathering and erosion.
4.3 Lithology and weathering
The type and composition of underlying rocks or soil
(lithology) play a significant role in landslide susceptibility.
Some lithological formations in NR are more prone to
weathering and erosion, leading to instability (e.g., alluvial,
quaternary deposits). Weathering processes weaken rock
and soil materials over time, making them more suscep-
tible to failure, particularly under the influence of external
triggers like precipitation.
4.4 Drainage density
Poor drainage exacerbates landslide susceptibility by increasing
water infiltration and reducing soil strength. High drainage
Figure 2: Geological map of the studied region.
Landslide susceptibility based on machine learning 5
density areas tend to have better natural drainage systems,
lowering the risk of landslides compared to regions with poor
drainage, where water accumulates and increases pore water
pressure.
4.5 Land-cover and vegetation
Vegetation acts as a stabilizing factor by reducing surface
erosion, intercepting rainfall, and binding soil particles
Figure 3: The landslide triggering factors for NR: (a) elevation, (b) aspect, (c) slope angle, (d) lithology, (e) drainage density, (f) distance to river, (g)
weathering, (h) land-cover, (i) precipitation, (j) NDVI, (k) distance to faults, (l) distance to roads, and (m) distance to the cities.
6Fei Teng et al.
together. Deforestation or land cover changes can increase
landslide susceptibility by removing this protective cover
and exposing slopes to erosion. Additionally, land cover
can affect surface runoffpatterns, further influencing
landslide occurrence.
Other factors such as distance to rivers, faults, roads,
and cities also influence landslide susceptibility indirectly.
Rivers and faults may act as weak zones or pathways for
water infiltration, while roads and urbanization can alter
natural drainage patterns and increase surface runoff,
thereby affecting slope stability. In landslide susceptibility
analysis, these triggering parameters are often integrated
into geospatial models using remote sensing and GIS techni-
ques to assess the likelihood of landslide occurrence in a given
area. By considering these factors collectively, researchers and
planners can identify high-risk areas, implement mitigation
measures, and develop strategies for sustainable land use
planning to reduce the impact of landslides on human lives
and infrastructure.
Remote sensing analysis, coupled with GIS, offers a
powerful tool for landslide susceptibility assessment in
the NR. By leveraging satellite imagery, researchers can
obtain high-resolution spatial data capturing various terrain
attributes and environmental factors influencing landslide
occurrence. Through image processing techniques, such as
spectral analysis, land cover, land use, slope characteristics,
and vegetation cover can be extracted and analyzed. GIS
allows for the integration of these remote sensing-derived
datasets with other spatial data layers, such as topography,
geology, and hydrology, to create comprehensive landslide
susceptibility models. By overlaying these layers and
applying statistical or machine learning algorithms, such
as logistic regression or random forest, researchers can
identify areas at high risk of landslide occurrence based
on the spatial distribution and interaction of contributing
factors. Moreover, satellite imagery enables temporal ana-
lysis, facilitating the monitoring of land cover changes, vege-
tation dynamics, and slope instability over time, which is
essential for assessing long-term landslide susceptibility
trends and guiding effective land management strategies
in the NR. For example, the normalized difference vegeta-
tion index (NDVI) is used as a vegetation index in mapping
which is a remote sensing index used to measure the health
and density of vegetation cover in an NR. The NDVI index is
based on the principle that healthy vegetation absorbs
visible light and reflects near-infrared light. NDVI values
range from −1to+1, with higher values indicating denser
and healthier vegetation cover [37]. NDVI is calculated using
the following formula [38]:
[][]=− +NDVI NIR Red / NIR Red ,
(1)
where NIR is the near-infrared band reflectance and Red is
the red band reflectance [38].
5 Methods
5.1 Comparative models implementation
The presented study used comparative predictive mod-
eling procedures for susceptibility assessments for land-
slides in NR. The SVM, MLP, and DT were selected for the
susceptibility analysis. Provided results were compared
and controlled by using well-known verification models.
SVM is a supervised learning algorithm that is particularly
effective when dealing with high-dimensional and non-
linear datasets, making it a popular choice [39] for land-
slide susceptibility assessments. The following steps are
demonstrating the predictive SVM-based model implemen-
tation in this article:
Step 1: Data preprocessing is used to gather the necessary
data and preprocess it. This involves cleaning and formatting
the data, as well as converting categorical variables into
numerical values. Additionally, the data need to be normal-
ized to ensure that all input variables are on the same scale.
Step 2: Feature selection which is used to select the
relevant triggering factors that influence landslides.
Step 3: Model training after input variables have been
selected (triggering factors); the next step is to train the
SVM model using a subset of the data.
Step 4: After training the model, it is essential to test its
performance on a separate subset of the data that was not
used for training.
Step 5: Prediction and mapping which the trained SVM
model is used to predict the probability of landslides in NR
based on input triggering factors and recorded historical
landslides.
MLP is a type of ANN that is commonly used for super-
vised learning tasks such as classification and regression. It
consists of one or more hidden layers of neurons, each of
which is connected to the input layer and the output layer
[40–48]. Implementing an MLP for a landslide suscept-
ibility assessment task involves several steps that can be
categorized into the following steps for NR:
–Steps 1 and 2: Same as steps 1 and 2 in SVM modeling,
–Step 3: Train-test split which splits the preprocessed
dataset into training and testing datasets.
–Step 4: Define the number of hidden layers, the number
of neurons in each layer, and the activation function
used in each neuron.
Landslide susceptibility based on machine learning 7
–Step 5: Compile the MLP model by specifying the loss func-
tion, optimizer, and metrics to be used during training.
–Step 6: Train the MLP model using the training dataset.
This involves feeding the training data into the MLP
model and updating the weights of the network through
backpropagation.
–Steps 7 and 8: Same as steps 4 and 5 in SVM modeling.
DT is a popular machine learning algorithm used for
classification and regression tasks [49]. It is a tree-like
model where each node represents a feature or attribute,
each branch represents a decision rule, and each leaf node
represents the outcome of the decision process [39,40].
Implementing a DT for a landslide susceptibility assess-
ment task involves several steps which can be categorized
into the following steps for NR:
–Steps 1 and 2: Same as steps 1 and 2 in SVM and MLP
modeling.
–Step 3: Same as step 3 in MLP modeling.
–Step 4: Define the criterion to be used for splitting the
data, such as entropy or Gini index.
–Step 5: Fit the DT model using the training dataset. This
involves recursively splitting the dataset into smaller sub-
sets based on the chosen criterion and hyperparameters.
–Steps 6 and 7: Same as steps 4 and 5 in SVM modeling or
steps 7 and 8 in MLP modeling.
Table 1 provides the hyperparameters used in this
article for each of the mentioned machine learning classi-
fiers for susceptibility analysis of landslides in NR. Hyper-
parameters are parameters that are not learned from the
data during model training but must be set before the
training process begins [51–54]. They are often used to
control the complexity of the model, regularize the model,
and avoid overfitting.
It is important to note that the effectiveness of SVM,
MLP, or DT for landslide susceptibility assessment depends
on several parameters including the quality and quantity
of the input data, the choice of feature selection and clas-
sification algorithms, and the inherent variability and
uncertainty in landslide processes. Therefore, it is essential
to carefully evaluate the model’s performance and limita-
tions before using it for decision-making purposes.
5.2 Data preparations
Twelve recorded historical landslides in NR are selected
for prone area identification. The aforementioned factors
are classified as triggering factors, as they render an area
vulnerable to movement without directly causing a land-
slide. Table 2 provides the main r resource of data pre-
pared for this study. Before these data can be used in
susceptibility modeling, they could be subject to multicol-
linearity and correlated variables. The multicollinearity is
a phenomenon in which one predictor variable in a regres-
sion model can be predicted linearly from others. To test
for multicollinearity, variance inflation factors (VIF) are
commonly used [10]. The VIF value of more than 5 indicates
potential multicollinearity. In this study, all triggering fac-
tors produced VIF values less than 2.25 are presented in
Table 2.
Therefore, these factors are responsible for the preva-
lence of landslides in NR where relevant data can be
obtained from available sources, historical development
background, and field studies. On the other hand, trig-
gering factors, such as rainfall and earthquakes, initiate
landslides by destabilizing the slope and converting it
from a marginally stable to an actively unstable state.
Generating an accurate landslide inventory map is crucial
for establishing the correlation between the 57 historical
landslide distributions and triggering factors in the NR. To
achieve this, extensive field surveys and observations were
conducted in the study area to produce a comprehensive
and dependable landslide inventory map. A digital eleva-
tion model (DEM), with a resolution of ±30 m, was used to
provide morphological maps in the county level. The DEM
data were used to prepare elevation, aspect, and slope
angle rasterized information layers in a GIS environment.
Using geo-units and ground data [34,35] helped to produce
the land cover, lithology, distance to faults, and raster
layers as well. In this study, a Landsat TM8 and ETM +
satellite image was used to provide weathering, vegetation,
distance to roads, and distance to the cities, drainage den-
sity, distance to river, relative information accordingly.
Figure 3 provides the rasterized maps of triggering factors
in GIS.
Table 1: The applied predictive models’hyperparameters
Classifiers Hyperparameters Elements
SVM Kernels Kernel = “poly”; degree = 2
Cvalue C= 100; Epsilon = 0.1
MLP Hidden layers’size Activation = “relu”
Learning rate Optimization = rmsprop
Optimization Loss_function = “mse”
Metrics = “mae”,“rmse”
DT Max depth Criterion = “gini”; Max_depth = 5
Random state Ccp_alpha = 0.0
Min_samples_leaf = 1
Random_state = 100
8Fei Teng et al.
Rasterizing triggering factors in ArcGIS involves con-
verting vector data representing factors such as elevation,
slope angle, land cover, and precipitation into raster format,
crucial for cohesive analysis. After gathering relevant vector
datasets, each is rasterized individually, assigning cell values
based on their attributes. Ensuring consistent spatial resolu-
tion and alignment among raster layers is essential for
seamless integration into a unified analysis framework.
Integration of rasterized layers entails overlaying them
using GIS operations, facilitating the development of land-
slide susceptibility maps. Various modeling techniques,
including statistical methods and machine learning algo-
rithms, leverage these integrated layers to assign suscept-
ibility values to different areas based on their risk profiles.
Validation of susceptibility maps with historical data or field
surveys ensures accuracy, providing insights into spatial dis-
tribution and the significance of triggering factors. Overall,
rasterizing triggering factors and integrating them into GIS-
based landslide susceptibility mapping workflows enable
comprehensive hazard assessments. This approach supports
informed decision-making for land use planning and risk
mitigation strategies, aiding in the identification of land-
slide-prone areas and understanding the factors driving sus-
ceptibility in the study region.
5.3 Models implementations and
verifications
After providing the database with undertaking
various triggering factors data as well as recorded histor-
ical landslides for NR. These data were used as primary
datasets for predictive modeling via SVM, MLP, and DT
classifiers. Models were trained with a training set and
tested with a testing set. These sets are provided by
random division of primary datasets of input data. The
training set was contain 70% of main database and testing
set is contain remained 30%. The ROC curve is commonly
used to evaluate the performance of predictive models. It is
a graph that plots the true positive rate (TPR) against the
false positive rate (FPR), which provides a measure of the
model’s overall accuracy and its ability to discriminate
between positive and negative cases. The ROC curve is a
useful tool for evaluating the performance of binary clas-
sification models, where the output variable can take on
one of two possible values (e.g., positive or negative). The
curve is created by varying the discrimination threshold of
the model and plotting the TPR against the FPR for each
threshold value. The TPR represents the proportion of true
positive cases correctly identified by the model, while the
FPR represents the proportion of false positive cases incor-
rectly identified by the model. In this research, the ROC
curve was employed as a validation tool to compare the
performance of different models for estimated overall
accuracy by AUC rates in preparing landslide susceptibility
maps. The AUC is often used as a summary statistic of the
model’s performance. A perfect classifier has an AUC of 1,
indicating that it can perfectly distinguish between positive
and negative cases. A random classifier has an AUC of 0.5,
indicating that its performance is no better than chance.
Comparative verification using the ROC curve involves
comparing the AUC values of different models to determine
which one performs better. A model with a higher AUC
value is generally considered to be more accurate and
reliable. In landslide susceptibility assessments, the ROC
Table 2: The landslide triggering factors information utilized in this work
Class Triggering factors Resolution (m) Data source VIF index
Morphologic Elevation ±30 DEM 1.25
Slope aspect ±30 DEM 1.27
Slope angle ±30 DEM 1.29
Geologic Lithology ±30 Geological data 1.96
Drainage density ±30 Landsat TM, IWRM
*
1.64
Distance to river ±30 Landsat TM, IWRM
*
1.54
Weathering ±30 Geological data 1.02
Land cover ±30 Geological data 1.02
Climatologic Precipitation ±30 Landsat TM, IMO
†
1.16
NDVI ±30 Landsat TM, IMO
†
1.12
Seismicity Distance to faults ±30 Seismic data 1.81
Human-activity Distance to roads ±30 DEM, Google Map 2.25
Distance to the cities ±30 DEM, Google Map 2.17
*
Iran Water Resources Management Company (IWRM).
†
Iran Meteorological Organization (IMO).
Landslide susceptibility based on machine learning 9
curve can be used to evaluate the performance of different
models in predicting the occurrence of landslides. By com-
paring the AUC values of different models, researchers
can identify the most effective model and determine the
range of discrimination thresholds that yield the best per-
formance. This information can be used to improve the
accuracy and reliability of landslide susceptibility assess-
ments and inform land-use planning and decision-making
processes.
6 Results
Applied machine learning models were comparatively used
toprovidemoreaccurateunderstandingregardingwhich
model can provide more reliable data for susceptibility ana-
lysis of landslide probability in the studied region. To this end,
various methods were used to identify and gather informa-
tion about triggering factors. These data were enriched by 57
recorded cases of large-scale landslides in NR. The identified
triggering factors can be classified as various elements. The
triggeringfactorsaswellasrecordedhistoricallandsides’
spatial location and magnitude were rasterized and entered
into the GIS environment as information layers; these layers
are normalized and utilized to prepare landslide suscept-
ibility maps for NR based on each target prediction proce-
dure. The prediction procedures are described in Section 6 of
this article properly. The results of the landslide susceptibility
mapping for NR are illustrated in Figure 4. According to this
figure,eachMLP,SVM,andDTclassifier provide suscept-
ibility maps for the studied region with some differences.
By comparatively looking at the maps, it will appear that
the SVM and MLP are nearly close to each other. A general
overview of the provided susceptibility maps can indicate
that the main high-risk areas are located on the west side
of the NR. These areas are related to the main faults and
seismic activities in the NR, especially the south-west region.
Most of the recorded historical landslides are identified in
that section.
Once the model is trained and validated, it is used to
predict landslide susceptibility for the entire study area.
The output of the model typically consists of continuous
probability values representing the likelihood of landslide
occurrence. These probabilities are then classified into five
susceptibility levels –very high, high, moderate, low, and
Figure 4: Comparatively prepared landslide susceptibility maps for NR: (a) MLP, (b) SVM, and (c) DT.
10 Fei Teng et al.
very low –based on predefined thresholds or decision
rules. This classification scheme enables the identification
of areas with varying degrees of susceptibility to landslides.
For example, in landslide susceptibility analysis, areas may
be classified into five classes, ranging from very low suscept-
ibility (indicating minimal risk of landslides) to very high
susceptibility (indicating a high likelihood of landslides).
The assignment of areas to susceptibility classes is typically
based on the analysis of various contributing triggering
factors and historical landslide occurrences. Susceptibility
classes provide valuable insights into the spatial distribution
of risk within a study area, enabling stakeholders to prioritize
mitigation efforts, implement land use planning measures,
and make informed decisions to reduce the impact of hazards
on communities and infrastructure.
In this study, the MLP, SVM, and DT models are
employed in landslide susceptibility analysis to define sus-
ceptibility classes and assess the likelihood of landslide
occurrence in NR. In this case, predictive models are
trained using historical landslide data and triggering fac-
tors. The models learns the complex relationships between
these factors and landslide occurrences to predict the like-
lihood of landslides in different areas. By analyzing the
output probabilities generated by the MLP model, suscept-
ibility classes can be defined, ranging from low to high
susceptibility, based on predefined thresholds. The SVM
model aims to find the hyperplane that best separates
the classes in the feature space. The distance between the
hyperplane and the data points, known as the margin, is
maximized to ensure robust classification. Based on the pre-
dicted probabilities or distances from the hyperplane, sus-
ceptibility classes can be defined, providing insights into the
spatial distribution of landslide risk. Additionally, DT models
recursively partition the feature space into subsets based on
the most discriminative features, such as slope, aspect, and
soil type. Each node in the DT represents a decision based on
aspecific feature, leading to the classification of areas into
different susceptibility classes. By analyzing the decision
paths and criteria used by the DT model, susceptibility classes
can be defined based on the resulting tree structure and the
likelihood of landslide occurrence associated with different
combinations of environmental factors.
The relationship between recorded historical land-
slides and their distance to the high and very high suscep-
tible zones is considerable. In such circumstances, it can be
stated that the MLP and SVM models are more accurate
than DT. Also, the higher frequency ratios of occurred land-
slide recorded to the close to the recorded landslides by the
SVM and MLP are need to be notified. ROC curve was used
to verify the predictive models’performances. To evaluate
the overall accuracy of the landslide susceptibility maps
and the binary classification model used to create it, ROC
verification can be used efficiently. ROC analysis involves
plotting the TPR (sensitivity) against the FPR (1 −specificity)
for different threshold values. This creates a curve that repre-
sents the performance of the model at different levels of
sensitivity and specificity. The AUC can be used as a measure
of the accuracy of the model. A higher AUC indicates better
performance, with a value of 1 indicating perfect accuracy. By
Figure 5: ROC analysis curve obtained for utilized predictive models.
Table 3: Landslide triggering factors
Class Triggering
factors
SVM
weight
MLP
weight
DT weight
Morphologic Elevation 4 5 5
Slope aspect 5 4 4
Slope angle 3 3 4
Geologic Lithology 5 4 5
Drainage
density
353
Distance to
river
543
Weathering 3 5 3
Land cover 5 4 4
Climatologic Precipitation 5 4 5
NDVI 4 5 3
Seismicity Distance to
faults
345
Human activity Distance to
roads
344
Distance to the
cities
345
Landslide susceptibility based on machine learning 11
comparing the AUC of different models or different input
parameters, it is possible to determine which factors are
most important in predicting landslide susceptibility and
which model is most accurate. Figure 5 presents the ROC
curve analysis results for MLP, SVM, and DT predictive
models. According to this figure, the MLP model reached
the highest accuracy rate than SVM and DT. The models’
weight indexes are provided in Tables 3 and 4.
In accordance with ROC results, it can be stated that the
MLP model with an overall accuracy of 0.883 is classified in the
highest rank for landslide susceptibility mapping. The SVM
model reached 0.842 as overall accuracy, which is located in
the second rank. DT predictive model provides the lowest
overall accuracy among other predictive classifiers.
7 Discussion
The scientific study explores the efficacy of machine learning
models in predicting landslide susceptibility within a region
denoted as NR. By integrating various factors along with his-
torical landslide data, the researchers created susceptibility
maps using different predictive models like MLP, SVM, and
DT. These factors were transformed into rasterized layers and
processed within a GIS environment, forming the basis for
the analysis. The resulting susceptibility maps exhibited dis-
tinct variations among the models employed. Notably, high-
risk areas were concentrated in the western part of NR,
aligning with main faults and seismic activity zones, which
correlated strongly with the occurrence of historical landslides.
The comparative analysis of the generated maps indicated that
MLP and SVM models demonstrated closer predictions in com-
parison to the DT model.
To evaluate the models’performances, the researchers
utilized ROC analysis, plotting sensitivity against 1 −speci-
ficity for different threshold values. This allowed for the
calculation of the AUC, a measure of model accuracy. The
findings revealed that MLP achieved the highest accuracy
with an AUC of 0.883, followed by SVM with an AUC of
0.842, while DT exhibited the least accuracy among the
models. In conclusion, the study highlighted that MLP
and SVM models showcased superior accuracy in pre-
dicting landslide susceptibility compared to DT. These find-
ings provide crucial insights for assessing and managing
landslide risks within the NR region, offering valuable gui-
dance for future mitigation strategies and risk manage-
ment protocols.
8 Conclusion
In conclusion, the landslide susceptibility analysis using
MLP, SVM, and DT algorithms has shown promising results
in predicting the likelihood of landslides in the studied
region (NR). Each of the three models has its own advan-
tages and disadvantages, and the choice of which model to
use will depend on the specific requirements of the study
and the characteristics of the area being analyzed. In the
studied region, the results of the overall accuracy analysis
indicated that the MLP reached the highest accuracy for
susceptibility analysis and mapping of landslide. The MLP
model has shown high accuracy in predicting landslide
susceptibility, but it can be computationally intensive
and requires a large amount of data for training. The
presented study used 70% of the primary database to
train the models (i.e., MLP, SVM, and DT) and the results
were tested for remaining 30% used to validate the pre-
diction process. The close prediction regarding SVM and
MLP might be related that the SVM is sensitive to the
choice of kernel function and hyperparameters, the
“poly”kernel function used properly in the landslide
susceptibility analysis for NR. The DT model is simple
to implement and interpret, but it is not as accurate as
the other two models. Nevertheless, the landslide sus-
ceptibility mapping for NR using these models has the
potential to assist in identifying areas that are at risk of
Table 4: Pixel comparison for output susceptibility models
Susceptibility class MLP model SVM model DT model
Pixel rate Percentage Pixel rate Percentage Pixel rate Percentage
Very low 359 12 385 10 218 8
Low 1,425 35 1,906 37 1,639 26
Moderate 1,247 21 1,054 22 2,159 35
High 963 24 636 19 756 21
Very high 165 8 105 12 210 10
High and Very high 1,128 32 741 31 966 31
12 Fei Teng et al.
landslides and guiding mitigation efforts to reduce the
impact of landslides on human society and the environ-
ment. Further research is needed to improve the accu-
racy of these models and to develop more effective
approaches to landslide susceptibility mapping.
Acknowledgments: The authors would like to thank the
anonymous reviewers for providing invaluable review
comments and recommendations for improving the scien-
tific level of the article.
Funding information: This research was funded by the Key
Improvement Projects of Guangdong Province (grant No.
2022ZDJS048), the Shaoguan Science and Technology Plan
Projects (grant No. SZ2022KJ06 and 220607154531533), and
the Science and Technology projects of Education Government
in Jiangxi province (grant No. GJJ209406, GJJ218505, and
GJJ218504) and the National Nature Sciences Foundation of
China (grant No. 42250410321).
Author contributions: Fei Teng, Yican Li and Subin Qian:
Methodology, Conceptualization, Formal analysis, Investigation,
Software, Data curation, Visualization, Writing –original
draft. Yimin Mao and Yaser A. Nanehkaran: Methodology,
Validation, Conceptualization, Investigation, Software,
Supervision, Writing –review & editing.
Conflict of interest: The authors declare that they have no
conflicts of interest to report regarding the present study.
Data availability statement: All required data and infor-
mation is available within paper.
References
[1] Eker AM, Dikmen M, Cambazoğlu S, Düzgün ŞH, Akgün H.
Evaluation and comparison of landslide susceptibility mapping
methods: a case study for the Ulus district, Bartın, northern Turkey.
Int J Geogr Inf Sci. 2015;29(1):132–58.
[2] Harrison JF, Chang CH, Liu CC. Identification of inventory-based
susceptibility models for assessing landslide probability: a case
study of the Gaoping River Basin, Taiwan. Geomat Nat Hazards
Risk. 2017;8(2):1730–51.
[3] Shano L, Raghuvanshi TK, Meten M. Landslide susceptibility eva-
luation and hazard zonation techniques–a review. Geoenviron
Disasters. 2020;7(1):1–19.
[4] Azarafza M, Ghazifard A, Akgün H, Asghari-Kaljahi E. Landslide
susceptibility assessment of South Pars Special Zone, southwest
Iran. Env Earth Sci. 2018;77:805.
[5] Varnes DJ. Slope movement types and processes. Washington:
Landslide Analysis and Control, Transportation Research Board,
National Academy Sciences; 1978.
[6] Highland LM, Bobrowsky P. The landslide handbook –a guide to
understanding landslides. Circular 1325. Reston, Virginia: US
Geological Survey; 2008.
[7] Nikoobakht S, Azarafza M, Akgün H, Derakhshani R. Landslide
susceptibility assessment by using convolutional neural network.
Appl Sci. 2022;12(12):5992.
[8] Bien TX, Truyen PT, Phong TV, Nguyen DD, Amiri M, Costache R,
et al. Landslide susceptibility mapping at sin Ho, Lai Chau province,
Vietnam using ensemble models based on fuzzy unordered rules
induction algorithm. Geocarto Int. 2022;37(27):17777–98.
[9] Sameen MI, Pradhan B, Lee S. Application of convolutional neural
networks featuring Bayesian optimization for landslide suscept-
ibility assessment. Catena. 2020;186:104249.
[10] Azarafza M, Azarafza M, Akgün H, Atkinson PM, Derakhshani R.
Deep learning-based landslide susceptibility mapping. Sci Rep.
2021;11(1):24112.
[11] Yong C, Jinlong D, Fei G, Bin T, Tao Z, Hao F, et al. Review of
landslide susceptibility assessment based on knowledge mapping.
Stoch Env Res Risk Ass. 2022;36(9):2399–417.
[12] Chen X, Chen W. GIS-based landslide susceptibility assessment
using optimized hybrid machine learning methods. Catena.
2021;196:104833.
[13] Nanehkaran YA, Mao Y, Azarafza M, Kockar MK, Zhu HH. Fuzzy-
based multiple decision method for landslide susceptibility and
hazard assessment: A case study of Tabriz. Iran Geomech Eng.
2021;24(5):407–18.
[14] Ercanoglu M, Gokceoglu C. Use of fuzzy relations to produce
landslide susceptibility map of a landslide prone area (West Black
Sea Region, Turkey). Eng Geol. 2004;75(3–4):229–50.
[15] Nhu VH, Hoang ND, Nguyen H, Ngo PTT, Bui TT, Hoa PV, et al.
Effectiveness assessment of Keras based deep learning with
different robust optimization algorithms for shallow
landslide susceptibility mapping at tropical area. Catena.
2020;188:104458.
[16] MarjanovićM, KovačevićM, Bajat B, Voženílek V. Landslide sus-
ceptibility assessment using SVM machine learning algorithm. Eng
Geol. 2011;123(3):225–34.
[17] Huang Y, Zhao L. Review on landslide susceptibility mapping using
support vector machines. Catena. 2018;165:520–9.
[18] Huang F, Cao Z, Jiang SH, Zhou C, Huang J, Guo Z. Landslide sus-
ceptibility prediction based on a semi-supervised multiple-layer
perceptron model. Landslides. 2020;17:2919–30.
[19] Li D, Huang F, Yan L, Cao Z, Chen J, Ye Z. Landslide susceptibility
prediction using particle-swarm-optimized multilayer perceptron:
Comparisons with multilayer-perceptron-only, bp neural network,
and information value models. Appl Sci. 2019;9(18):3664.
[20] Yeon YK, Han JG, Ryu KH. Landslide susceptibility mapping in Injae,
Korea, using a decision tree. Eng Geol. 2010;116(3–4):274–83.
[21] Lee S, Ryu JH, Won JS, Park HJ. Determination and application of the
weights for landslide susceptibility mapping using an artificial
neural network. Eng Geol. 2004;71:289–302.
[22] Ermini L, Catani F, Casagli N. Artificial neural networks applied to
landslide susceptibility assessment. Geomorphology.
2005;66:327–43.
[23] Kanungo DP, Sarkar S, Sharma S. Combining neural network with
fuzzy, certainty factor and likelihood ratio concepts for spatial
prediction of landslides. Nat Hazards. 2011;59:1491–512.
[24] Oh HJ, Pradhan B. Application of a neuro-fuzzy model to landslide-
susceptibility mapping for shallow landslides in a tropical hilly area.
Comput Geosci. 2011;37:1264–76.
Landslide susceptibility based on machine learning 13
[25] Quan HC, Lee BG. GIS-based landslide susceptibility mapping using
analytic hierarchy process and artificial neural network in Jeju
(Korea). KSCE J Civ Eng. 2012;16:1258–66.
[26] Park S, Choi C, Kim B, Kim J. Landslide susceptibility mapping using
frequency ratio, analytic hierarchy process, logistic regression, and
artificial neural network methods at the Inje area, Korea. Env Earth
Sci. 2013;68:1443–64.
[27] Nourani V, Pradhan B, Ghaffari H, SharifiSS. Landslide suscept-
ibility mapping at Zonouz Plain, Iran using genetic programming
and comparison with frequency ratio, logistic regression, and
artificial neural network models. Nat Hazards. 2014;71:523–47.
[28] Liu Y, Wu L. Geological disaster recognition on optical remote
sensing images using deep learning. Procedia Comput Sci.
2016;91:566–75.
[29] Xiao L, Zhang Y, Peng G. Landslide susceptibility assessment using
integrated deep learning algorithm along the China-Nepal
highway. Sensors. 2018;18:4436.
[30] Ortiz JAV, Martínez-Graña AM. A neural network model applied to
landslide susceptibility analysis (Capitanejo, Colombia). Geomat
Nat Hazards Risk. 2018;9:1106–28.
[31] Ghorbanzadeh O, Blaschke T, Gholamnia K, Meena SR, Tiede D,
Aryal J. Evaluation of different machine learning methods and
deep-learning convolutional neural networks for landslide detec-
tion. Remote Sens. 2019;11:196.
[32] Mutlu B, Nefeslioglu HA, Sezer EA, Akcayol MA, Gokceoglu C. An
experimental research on the use of recurrent neural networks in
landslide susceptibility mapping. ISPRS Int J Geo-Inf. 2019;8:578.
[33] Sarkar S, Kanungo DP. An integrated approach for landslide sus-
ceptibility mapping using remote sensing and GIS. Photogram Eng
Rem Sens. 2004;70(5):617–25.
[34] Aghanabati A. Geology of Iran. Geological Survey & Mineral
Explorations of Iran Press; 2009.
[35] Geological Survey of Iran, Geological data and maps for Naqadeh
region. Geological Survey & Mineral Explorations of Iran
Press; 2009.
[36] Tahroudi MN, Ramezani Y, De Michele C, Mirabbasi R. Analyzing the
conditional behavior of rainfall deficiency and groundwater level
deficiency signatures by using copula functions. Hydrol Res.
2020;51(6):1332–48.
[37] Pradhan B, Sezer EA, Gokceoglu C, Buchroithner MF. Landslide
susceptibility mapping by neuro-fuzzy approach in a landslide-
prone area (Cameron Highlands, Malaysia). IEEE Trans Geosci Rem
Sens. 2010;48(12):4164–77.
[38] Huang S, Tang L, Hupy JP, Wang Y, Shao G. A commentary review
on the use of normalized difference vegetation index (NDVI) in the
era of popular remote sensing. J Forestry Res. 2021;32(1):1–6.
[39] Aggarwal CC. Neural networks and deep learning: A textbook.
Springer; 2018.
[40] Müller AC, Guido S. Introduction to machine learning with Python:
A guide for data scientists. O’Reilly Media; 2016.
[41] Mao Y, Mwakapesa DS, Wang G, Nanehkaran YA, Zhang M.
Landslide susceptibility modelling based on AHC-OLID clustering
algorithm. Adv Space Res. 2021;68(1):301–16.
[42] Yimin M, Yican L, Simon Mwakapesa D, Genglong W,
Nanehkaran YA, Asim Khan M, et al. Innovative landslide suscept-
ibility mapping portrayed by CA-AQD and K-means clustering
algorithms. Adv Civ Eng. 2021;2021:1–17.
[43] Mao Y, Mwakapesa DS, Li YC, Xu KB, Nanehkaran YA, Zhang MS.
Assessment of landslide susceptibility using DBSCAN-AHD and LD-
EV methods. J Mt Sci. 2022;19(1):184–97.
[44] Huang F, Xiong H, Jiang SH, Yao C, Fan X, Catani F, et al. Modelling
landslide susceptibility prediction: A review and construction of semi-
supervised imbalanced theory. Earth-Sci Rev. 2024;250:104700.
[45] Huang F, Cao Z, Guo J, Jiang SH, Li S, Guo Z. Comparisons of
heuristic, general statistical and machine learning models for
landslide susceptibility prediction and mapping. Catena.
2020;191:104580.
[46] Huang F, Zhang J, Zhou C, Wang Y, Huang J, Zhu L. A deep learning
algorithm using a fully connected sparse autoencoder neural network
for landslide susceptibility prediction. Landslides. 2020;17:217–29.
[47] Huang F, Xiong H, Yao C, Catani F, Zhou C, Huang J. Uncertainties of
landslide susceptibility prediction considering different landslide
types. J Rock Mech Geotech Eng. 2023;15(11):2954–72.
[48] Huang F, Teng Z, Yao C, Jiang SH, Catani F, Chen W, et al.
Uncertainties of landslide susceptibility prediction: influences of
random errors in landslide conditioning factors and errors reduc-
tion by low pass filter method. J Rock Mech Geotech Eng.
2024;16(1):213–30.
[49] Daviran M, Shamekhi M, Ghezelbash R, Maghsoudi A. Landslide
susceptibility prediction using artificial neural networks, SVMs and
random forest: hyperparameters tuning by genetic optimization
algorithm. Int J Environ Sci Technol. 2023;20(1):259–76.
[50] Bui DT, Tsangaratos P, Nguyen VT, Van Liem N, Trinh PT.
Comparing the prediction performance of a Deep Learning Neural
Network model with conventional machine learning models in
landslide susceptibility assessment. Catena. 2020;188:104426.
[51] Adnan MSG, Rahman MS, Ahmed N, Ahmed B, Rabbi MF,
Rahman RM. Improving spatial agreement in machine learning-
based landslide susceptibility mapping. Remote Sens.
2020;12(20):3347.
[52] Ali SA, Parvin F, Pham QB, Khedher KM, Dehbozorgi M, Rabby YW,
et al. An ensemble random forest tree with SVM, ANN, NBT, and
LMT for landslide susceptibility mapping in the Rangit River
watershed, India. Nat Hazards. 2022;113(3):1601–33.
[53] Ado M, Amitab K, Maji AK, Jasińska E, Gono R, Leonowicz Z, et al.
Landslide susceptibility mapping using machine learning: A litera-
ture survey. Remote Sens. 2022;14(13):3029.
[54] Liu Q, Tang A, Huang D. Exploring the uncertainty of landslide
susceptibility assessment caused by the number of non–landslides.
Catena. 2023;227:107109.
14 Fei Teng et al.
