You are on page 1of 15

Geoscience Frontiers 12 (2021) 505–519

H O S T E D BY Contents lists available at ScienceDirect

Geoscience Frontiers
journal homepage: www.elsevier.com/locate/gsf

Research Paper

Evaluation of deep learning algorithms for national scale landslide


susceptibility mapping of Iran
Phuong Thao Thi Ngo a, Mahdi Panahi b, c, Khabat Khosravi d, Omid Ghorbanzadeh e,
Narges Kariminejad f, Artemi Cerda g, Saro Lee c, h, *
a
Institute of Research and Development, Duy Tan University, Da Nang, 550000, Viet Nam
b
Division of Science Education, Kangwon National University, Chuncheon-si, Gangwon-do, 24341, Republic of Korea
c
Geoscience Platform Division, Korea Institute of Geoscience and Mineral Resources (KIGAM), 124, Gwahak-ro Yuseong-gu, Daejeon, 34132, Republic of Korea
d
Department of Watershed Management Engineering, Sari Agricultural Science and Natural Resources University, Sari, Iran
e
Department of Geoinformatics–Z_GIS, University of Salzburg, Salzburg, 5020, Austria
f
Department of Watershed and Arid Zone Management, Gorgan University of Agricultural Sciences and Natural Resources, Gorgan, 49189-434, Iran
g
Soil Erosion and Desertification Research Group, Department of Geography, University of Valencia, Valencia, Spain
h
Department of Geophysical Exploration, Korea University of Science and Technology, 217 Gajeong-ro Yuseong-gu, Daejeon, 34113, Republic of Korea

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

Handling Editor: E. Shaji The identification of landslide-prone areas is an essential step in landslide hazard assessment and mitigation of
landslide-related losses. In this study, we applied two novel deep learning algorithms, the recurrent neural
Keywords: network (RNN) and convolutional neural network (CNN), for national-scale landslide susceptibility mapping of
CNN Iran. We prepared a dataset comprising 4069 historical landslide locations and 11 conditioning factors (altitude,
RNN
slope degree, profile curvature, distance to river, aspect, plan curvature, distance to road, distance to fault,
Deep learning
rainfall, geology and land-sue) to construct a geospatial database and divided the data into the training and the
Landslide
Iran testing dataset. We then developed RNN and CNN algorithms to generate landslide susceptibility maps of Iran
using the training dataset. We calculated the receiver operating characteristic (ROC) curve and used the area
under the curve (AUC) for the quantitative evaluation of the landslide susceptibility maps using the testing
dataset. Better performance in both the training and testing phases was provided by the RNN algorithm (AUC ¼
0.88) than by the CNN algorithm (AUC ¼ 0.85). Finally, we calculated areas of susceptibility for each province
and found that 6% and 14% of the land area of Iran is very highly and highly susceptible to future landslide
events, respectively, with the highest susceptibility in Chaharmahal and Bakhtiari Province (33.8%). About 31%
of cities of Iran are located in areas with high and very high landslide susceptibility. The results of the present
study will be useful for the development of landslide hazard mitigation strategies.

1. Introduction Landslides are characterized by the downslope movement of rock,


soil, and debris due to gravity; they are classified based on their
Landslides cause extensive losses worldwide in terms of economic component material (e.g., rock, debris, soil, or mud) and movement type
and natural resources and human life (Confuorto et al., 2019). Landslides (e.g., slide, flow, or topple) (Nachappa et al., 2019). Landslide triggering
are among the most frequent and destructive natural hazards in moun- factors depend on the characteristics of the study area. Usually, they
tainous areas, damaging infrastructure components such as roads, include natural factors (e.g., heavy rainfall, earthquake, or rapid snow-
bridges, and power lines, as well as soils and vegetation, triggering melt) and/or human-made factors (e.g., road construction or irrigation)
intense soil erosion processes that can lead to badland development and (Wilde et al., 2018). Although landslides are typically considered natural
other forms of land degradation (Pourghasemi and Rahmati, 2018). hazards, their frequency is influenced by anthropogenic factors (Lima
Landslide mitigation is, therefore, an essential step toward achieving the et al., 2017). In recent years, surges in population growth, infrastructure
United Nations Sustainable Goals for Development. development, settlement expansion, and road construction have

* Corresponding author. Geoscience Platform Division, Korea Institute of Geoscience and Mineral Resources (KIGAM), 124, Gwahak-ro Yuseong-gu, Daejeon, 34132,
Republic of Korea
E-mail addresses: ngotphuongthao5@duytan.edu.vn (P.T. Thi Ngo), khabat.khosravi@gmail.com (K. Khosravi), leesaro@kigam.re.kr (S. Lee).
Peer-review under responsibility of China University of Geosciences (Beijing).

https://doi.org/10.1016/j.gsf.2020.06.013
Received 2 March 2020; Received in revised form 9 May 2020; Accepted 19 June 2020
Available online 7 August 2020
1674-9871/© 2020 China University of Geosciences (Beijing) and Peking University. Production and hosting by Elsevier B.V. This is an open access article under the
CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

increased the likelihood of landslide occurrence in mountainous regions algorithms such as RF, LMT and Baysian algorithms have a high pre-
in developing countries, substantially increasing landslide-related mor- diction power, selecting a proper algorithm for big data, like present
tality rates (Ghorbanzadeh et al., 2019a). This phenomenon has attracted study, is challenging and deep learning algorithms which developed to be
the attention of the global scientific community, leading to the publica- applied in big dataset and rarely applied so far, have to be performed and
tion of a large number of studies examining landslide risk assessment evaluated.
(Nachappa et al., 2019). Recently, deep learning (DL) models, particularly recurrent neural
In recent years, many landslides have occurred in Iran due to a network (RNN) and convolutional neural network (CNN) models, have
combination of intense tectonic activity and multiple factors, including been successfully used in a wide range of applications and be optimal for
meteorology, climate, vegetation cover, and anthropogenic activity (e.g., big data handling (Wang et al., 2020b). Like all DL models, RNN consists
poor watershed management) (Dehnavi et al., 2015). Landslide suscep- of layers, learnable parameters, and loss functions (Zaremba et al., 2014).
tibility mapping (LSM) is the first and most crucial step in landslide CNNs are distinguished from RNNs by having convolutional and pooling
hazard management and mitigation policy. Versatile and reliable LSM of layers (Krizhevsky et al., 2012), and by considering only the current
an area can be achieved by assessing the impact of relevant landslide input data, whereas RNNs consider current input data and previously
conditioning factors, resulting in the spatial distribution of landslide received inputs (Yin et al., 2017). CNNs have achieved good results in
occurrence probabilities, expressed as risk categories as well as applica- object detection (Guirado et al., 2017; Ghorbanzadeh et al., 2019a) and
tion of reliable and effective algorithm. Such maps can provide stake- semantic segmentation (Long et al., 2015). In contrast, RNNs show better
holders with insight into high landslide risk areas to guide land use performance in image recognition and characterization (Du et al., 2015),
management decisions. as well as the analysis of sequential data such as time series spatial data
LSM is often performed using different conditional factors and (Sauter et al., 2010). Although CNNs and RNNs have obtained acceptable
methodologies for small regions within a country (Guzzetti et al., 2012), results in various fields, their real efficiency on landslide modeling and
which is an inefficient process that prevents comparisons among regions LSM capabilities on a large scale or big data have not been explored (Can
(Chen et al., 2018). However, comprehensive and sustainable et al., 2019; Ghorbanzadeh and Blaschke, 2019; Ghorbanzadeh et al.,
national-scale landslide management requires the identification of 2019b). A few DL models have been applied for natural hazard suscep-
landslide-susceptible areas throughout the country, to enhance cooper- tibility mapping including flash floods (Bui et al., 2020a) and landslide
ation among local and national decision-makers, to cope with significant susceptibility mapping (Bui et al., 2020b; Wang et al., 2020a). They used
landslide events (Tehrany et al., 2015). For example, Iran is considered a different DL algorithms separately while their efficiency have not
landslide hot-spot because it has experienced many of the world’s most compared yet. To our knowledge, no study has applied RNN or CNN for
massive landslides, which have caused severe infrastructure damage and LSM in at national scale as well. In the present study, effectiveness of
high economic losses (i.e. saidmareh) (Feizizadeh et al., 2014; Dehnavi RNN and CNN algorithms examined for LSM at Iran’s national-scale. This
et al., 2015). Several studies have applied LSM to parts of Iran at different is the first study to apply and compare these DL models for LSM at a
regional and national scales and using a wide range of conditioning national scale; we, therefore, expect that it will represent a significant
factors such as land use, geology, stream and road networks, elevation, contribution to the scientific literature.
ground slope, aspect, and profile curvature (Trigila et al., 2013; Van
Westen, 2013; Dehnavi et al., 2015). However, accurate recognition of 2. Material and methods
Iran’s susceptible areas for landslide occurrences via reliable algorithm,
still remains as a big challenge. 2.1. Study area
Recent advancements in remote sensing (RS) and geographic infor-
mation systems (GIS) technology have significantly enhanced the effi- Iran (82 million inhabitants) extends from 25 N to 40 N and from
ciency of LSM and the reliability of its results. RS-derived data and GIS 44 E to 63 50 3000 E, with a total land area of 1.648 million km2 (Fig. 1).
spatial analysis tools are the foundation of LSM, and landslide condi- The land surface of Iran is generally mountainous in the north and west,
tioning factors and detailed landslide inventory data are essential for with plains and salty arid deserts in other regions. Slope and elevation
both knowledge-based and data-driven spatial modeling methods vary in a range from 0 to 60 and from 10 to 4500 m, respectively. The
(Ghorbanzadeh et al., 2018a; Piralilou et al., 2019). Several studies have highest elevations in Iran lie in the Alborz and Zagros mountain ranges,
implemented expert knowledge-based spatial modeling methods to which are extremely prone to landslides due to a combination of
generate natural hazard susceptibility maps; these include analytical geomorphological setting and climate. High-mountain landslides pose a
hierarchy process (AHP) (Feizizadeh et al., 2014), analytic network severe threat to public property and infrastructure in downslope areas.
process (ANP) (Neaupane and Piantanakulchai, 2006), interval pairwise Western and northern Iran are characterized by high humidity, with
comparison matrix (IPCM) (Ghorbanzadeh et al., 2018b), evidential mean annual rainfall ranging as 450–1800 mm, whereas southern and
belief function (EBF) (Bui et al., 2012; Nampak et al., 2014), and fuzzy eastern Iran have drier climates, with mean annual rainfall ranging as
logic models, as well as data-driven spatial modeling methods such as 100–300 mm (Iran Meteorological Organization data). The majority
support vector machine (SVM) (Chen et al., 2017), random forest (RF) lands in Iran covered by mountains (50%), followed by 25% desert and
(Chapi et al., 2017), logistic model tree (LMT) (Chen et al., 2019b), 25% productive plains (Dareshoori and Kasraian, 1998). Iran’s Geology
step-wise weight assessment ratio analysis (SWARA) (Pourghasemi et al., is very diverse, more than 558 lithological units (Dehnavi et al., 2015),
2019), Bayesian algorithm (Pham et al., 2019), logistic regression (LR) and it has very diverse tectonic features due to the location in the middle
(Felicísimo et al., 2013), and artificial neural network (ANN) models (Lee of Alpine–Himalayan orogenic belt (Dehnavi et al., 2015), more infor-
et al., 2004). Expertise-based models suffer from their dependency on mation can be found in Dehnavi et al. (2015).
expert opinion, which is a source of bias (Khosravi et al., 2018). SVM
models contain four kernels, which reduces bias by determining the most 2.2. Methodology
effective kernel during the modeling effort. ANN is the most widely used
algorithm; however, it has weak prediction power, especially when range In this study, the main steps followed to generate landslide suscep-
of testing dataset is out from training data. The adaptive neuro-fuzzy tibility maps included data collection (historical landslide locations and
inference system (ANFIS) is a hybrid of ANN and fuzzy logic; it was conditioning factors), assessing data accuracy and removing noisy data
developed to resolve issues with ANN modeling but has difficulty iden- with null prediction power, overlaying landslide locations with each
tifying membership function weights. Therefore, researchers continue to conditioning factor to attribute value extraction, landslide probability
seek new and more robust algorithms for application at different spatial index (LPI) prediction using CNN and RNN, LPI classification and put
scales (Chen et al., 2019a; Wang et al., 2019). Although data mining them in a five susceptibility classes (i.e. very low, low, moderate, high

506
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Fig. 1. (a) Map of Iran showing past and present landslide events; (b) provinces and (c) cities of Iran.

507
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

and very high), model validation, overlaying cities and provinces with 2.2.3. Landslide conditioning factor assessment
landslide zonation, and finally risk index calculation. A conceptual model A review of the literature suggested a large number of conditioning
of these steps is shown in Fig. 2. factors for LSM preparation and landslide hazard assessment. However,
only a few of these factors have predictive power and some are associated
2.2.1. Historical landslide inventory map with null effectiveness, which must be removed from the modeling
To determine the relationships between landslide occurrence and process to improve the predictive power of the applied algorithms. At the
various conditioning factors and to identify common landslide failure first stage, the selection of conditioning factors was made according to
mechanisms (Guzzetti et al., 2012; Pham et al., 2019), we first acquired literature review and data availability. Afterwards, all considered con-
historical landslide event data for the study area from different sources, ditioning factors are common between different types of landslides.
including Iranian water resources management company, Google Earth, Finally, we applied the most widely used information gain ranking filter
Forest Range and Watershed Management Organization and agricultural (IGRF) technique to evaluate the importance of each conditioning factor.
and natural resources management bureaus. We compiled these data as a IGRF value varies between zero (no effective) to 1 (most effective). Those
landslide inventory map of Iran, which comprised 4069 landslide event conditioning factors that do not add much information will have a lower
locations (Fig. 1). The most common type of landslide in Iran was the score and can be removed. The selected conditioning factors are
rotational slide. described in Fig. 3 and Table 1.

2.2.2. Training and testing dataset preparation 2.2.4. Model description


In artificial intelligence-based LSM, spatial prediction consists of bi- ANNs are loosely modeled on the human brain structure (Fig. 3). To
nary classification of data into two groups: landslide and non-landslide. process image data, ANNs require large amounts of input. For example,
We randomly generated non-landslide locations using the ArcGIS 10.2 an image represented by a 1000  1000 pixel matrix requires an input
software. We used 70% of the data in each group as a training dataset for layer of 1,000,000 neural nodes, with additional nodes for the hidden
model building, and the remaining 30% as a testing dataset for model layer. As the number of network layers increases, the total structure of
evaluation. The training and testing datasets were converted to a raster the network becomes more complicated, increasing the number of
format with a pixel size of 85 m  85 m (i.e. same as the digital elevation weight parameters and generally hindering the training procedure (Lee
model spatial resolution). Landslide and non-landslide locations were et al., 2017). To address this problem, Bengio et al., (1994) developed the
coded as 1 and 0, respectively (Chen et al., 2019a). Finally, the training CNN, which incorporates local connections among layers. Thus, CNNs
dataset was overlaid with 11 landslide conditioning factors for attribute consist of multiple neuronal layers usually including convolution, acti-
value extraction. vation, and pooling layers (Fig. 4) (LeCun and Bengio, 1995; LeCun et al.,

Fig. 2. Graphical representation of conceptual model.

508
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Fig. 3. Landslide conditioning factors: (a) altitude, (b) slope degree, (c) aspect, (d) plan curvature, (e) profile curvature, (f) distance to river, (g) distance to road, (h)
distance to fault, (i) rainfall, (j) geology, (k) land use.

1998). The convolution layer preserves relationships among pixels by ability of the network. For each new input, the output is determined and
learning image features from small squares of input data via a mathe- then fed back as the modified input. This operation is continued until a
matical operation with two inputs, such as an image matrix and a filter or constant output has been attained (Siegelmann and Sontag, 1991; Jaeger,
kernel. The activation layer consists of a nonlinear activation function 2002). The steps of this process are as follows:
appearing after the convolutional layer. The most common and effective X X
activation function is the rectified linear unit (ReLU) function (Nair and ai ðtÞ ¼ ϕji Xi ðtÞ þ σ ji hi ðt  1Þ; j ¼ 1; …; nH (1)
Hinton, 2010); intuitively, its gradients are always 0 s and 1 s. The
pooling layer reduces the number of parameters when images are too hi ðtÞ ¼ Fðai ðtÞÞ; j ¼ 1; …; nH (2)
large while retaining the essential information. In a deep neural network,
a fully connected simple multilayer perceptron performs discriminative X
bi ðtÞ ¼ aji hi ðtÞ; j ¼ 1; …; nJ (3)
learning to identify an object class. Numerous CNN structures have been
developed according to the data and image types, including ZFNet,  
VGGNet, GoogleNet, ResNet, LeNet-5, and AlexNet (Fang et al., 2019). Zi ðtÞ ¼ G bj ðtÞ ; j ¼ 1; …; nJ (4)
Comprehensive descriptions of the structure, initial training parameters,
and performance of these CNNs have been published previously. where ϕji, σ ji, and aji are weights. During the learning process, hidden
RNNs involve feedback loops from network output Z1(t), …, Znj to neurons receive the input vector X(t) to compute a modification to the
network inputs X(t). These loops significantly influence the learning network weights. F is a nonlinear transformation that occurs between

509
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Fig. 3. (continued).

neurons. The hidden neuron number nH is established by trial and error. 2.2.6. Map accuracy evaluation
In a simple RNN, neurons in the middle layer obtain input values from The models developed based on the training dataset; hence, results of
both input neurons and hidden neurons (Fig. 5). this section only show how the developed algorithms are fitted with the
dataset. Therefore, we evaluated the performance of the CNN and RNN
2.2.5. Landslide susceptibility map generation algorithms using the testing dataset, which was not used in the modeling
In this step, the entire area of Iran was converted to a raster format process. The testing dataset for evaluation purpose illustrates how well
with a pixel size of 85 m. Models trained using a CNN and an RNN were the model predicts landslide occurrences. We calculated a receiver
then evaluated using the testing dataset. Landslide susceptibility indices operating characteristic (ROC) curve for the testing dataset to determine
were calculated for each pixel in the study area using each model; these the prediction rate, and used the area under the curve (AUC) to quanti-
indices were then exported from the MATLAB programming environ- tatively assess model performance (Bui et al., 2019; Rahmati et al.,
ment into the ArcGIS 10.2 software. The indices were converted into 2020). The AUC varied between 0.5 and 1, with higher AUC values
landslide susceptibility maps, and areas were classified using the quantile indicating better performance. Yesilnacar (2005) classified performance
method into five categories: very low, low, moderate, high, and very high of each algorithm to 0.5–0.6, 0.6–0.7, 0.7–0.8, 0.8–0.9, and 0.9–1 which
landslide susceptibility. The quantile method is widely used for classifi- are responsible for poor, average, good, very good and excellent
cation and threshold value identification with skewed data (Akgun et al., performance.
2012).

510
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Fig. 3. (continued).

3. Results landslide areas more successfully than the CNN algorithm in both the
training and testing phases (Figs. 6c, e, and 7c, e, respectively), with an
3.1. Variable importance analysis error range of 0.2 to 2 for RNN and 0.4 to 0.4 for CNN in both phases.
RMSE values for the RNN algorithm during the training and testing
Our IGRF results showed that all considered input variables signifi- phases were 0.081 and 0.083, respectively, whereas those of the CNN
cantly affected the landslide susceptibility modeling output (Table 2). algorithm were 0.131 and 0.130, indicating better performance by the
Slope degree had the highest impact on landslide occurrence in Iran RNN algorithm. Standard deviation (StD) of error value in the training
(IGRF ¼ 0.821), followed by geology (0.683), land use (0.629), distance and testing phase was lower for the RNN algorithm (0.080; 0.083) than
from nearest fault (0.60), plan curvature (0.5431), rainfall (0.471), dis- for the CNN algorithm (0.131; 130), confirming the superior perfor-
tance from nearest road (0.439), altitude (0.425), profile curvature mance of the RNN model as it shows lower range of error.
(0.420), distance from nearest river (0.25), and aspect (0.08). The outputs of the CNN- and RNN-based models were exported and
analyzed using the ArcGIS software to generate landslide susceptibility
maps of Iran (Fig. 8). Both maps showed similar results, revealing very
3.2. Model results and map generation
high susceptibility to future landslide events in northern, northwestern,
and western Iran. A visual comparison of conditioning factors with LSM
In this study, we applied CNN- and RNN-based models with optimized
shows that very highly susceptible area have slope, altitude, and rainfall
parameters to identify landslide-prone areas in Iran. Relationships be-
values exceeding 20%, 2600 m, and 800 mm, respectively.
tween input and output variables determined by these models in the
Around 3.4% higher predictive power was observed for the RNN al-
training and testing phases are shown in Figs. 6 and 7, respectively.
gorithm (AUC ¼ 0.88) than for the CNN algorithm (0.85) using both the
Model performance was assessed in terms of root mean square error
training and testing datasets (Fig. 9). The RNN- and CNN-based models
(RMSE). The relationship between the target (past and present landslide
showed accuracy values of 88% and 85%, respectively.
event) and output (predicted landslide event) over time is shown in
Based on the model output of the RNN algorithm, 20% (283,620 km2)
Figs. 6a, b, and 7a, b. The RNN algorithm predicted landslides and non-

511
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Table 1 very high landslide risk (45%, 14 cities), followed by Lorestan (40%, ten
Input variables description. cities), Kohgiluyeh and Boyer-Ahmad (37.5%, six cities), Kurdistan
Conditioning Source Spatial Importance (32%, eight cities), Markazi (21.8%, seven cities), and Mazandaran
factor resolution (20.7%, 11 cities) (Fig. 13). However, larger percentages of cities with
Aspect DEM 80 m  80 Different amount of rainfall, soil very low landslide risk were detected in Yazad (66%), Tehran (39.7%),
m moisture due to different solar Sistan and Balouchestan (51.3%), Qom (50%), North Khorasan (44.2%),
radiation and vegetation cover and Boushehr (46.4%).
Slope DEM 80 m  80 Infiltration process, shear stress
m and gravity effect
Plan and profile DEM 80 m  80 Erosion process and control 4. Discussion
curvature m surface runoff
Altitude DEM 80 m  80 Has effect on rainfall and In landslide hot-spot regions, landslide prediction is necessary for
m vegetation cover as well as soil
decision-makers, engineers, and authorities to determine appropriate
depth
Rainfall National 80 m  80 Effect on soil moisture land use management, hazard mitigation, and watershed management
map m practices. The identification of areas with high or very high landslide
Distance from National 1:25,000 Cutting the slope hill is more susceptibility provides a tool for risk management and the development
roads map sensitive to landslide
of innovative responses to future landslides. Landslide prediction is a part
Distance from National 1:100,000 Effect on the soil moisture
rivers map of the United Nations Sustainable Development Goals 2030 perspective
Distance from National 1:100,000 Cause discontinuity of soil and and this paper will contribute to this achievement (Keesstra et al., 2016).
faults map rocks Many studies have examined landslide susceptibility in Iran using various
Geology National 1:100,000 Different formation has a different techniques including fuzzy ANP in West Azerbaijan Province (Alilou
map structure and permeability
et al., 2019), machine learning models in Kurdistan Province (Shirzadi
Land-use National 1:25,000 Vegetation root cause to more
map stability et al., 2019), and Mazandaran Province (Pourghasemi and Rahmati,
2018). However, worldwide increases in landslide frequency and in-
tensity are anticipated due to climate change and poor watershed
of the land area of Iran shows very high or high landslide susceptibility.
In contrast, 20% (283,620 km2), 24% (340,098 km2), and 36% (503,810
km2) of Iran’s land area shows moderate, low, or very low landslide
susceptibility, respectively (Fig. 10).
The nine provinces with the highest percentages of land area with
very high landslide susceptibility were Chaharmahal and Bakhtiari,
33.8% (5515 km2); Gilan, 33.3% (4599 km2); Kohgiluyeh and Boyer-
Ahmad, 33.2% (5149 km2); Lorestan, 31.3% (8833 km2); Kurdistan,
32.9% (9528 km2); Masandaran, 31% (7334 km2); Kermanshah, 20%
(5247 km2); West Azarbayjan, 19.7% (7173 km2); and Zanjan, 14%
(3053 km2) (Fig. 11). The provinces with the highest percentages of land
area with very low landslide susceptibility are Semnan, 66% (64,863
km2); Qom, 65% (7515 km2); Yazd, 55% (70,881 km2); South Khorasan,
48.7% (40,814 km2); Sinstan and Balouchestan, 46% (75,736 km2);
Khozestan, 43.2% (28,817 km2); Kerman, 42% (71,972 km2); and Isfa-
han, 40% (46,278 km2).
Overall, 12% (137), 19% (210), 20% (227), 19% (215), and 30%
(328) of all cities in Iran are located in areas showing very high, high,
moderate, low, or very low landslide susceptibility (Fig. 12). Cha-
harmahal and Bakhtiari Province, which has a large proportion of area
susceptible to landslides, also has the highest percentage of cities with Fig. 5. Recurrent neural network (RNN) architecture.

Fig. 4. Convolutional neural network (CNN) structure.

512
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Table 2
Importance of landslide conditioning factors.
Variable Slope Plan Profile Geology Land- Distance from Rainfall Distance from Aspect Distance from Altitude
degree curvature curvature use fault road river

IGRF 0.821 0.543 0.42 0.683 0.629 0.6 0.471 0.439 0.08 0.25 0.425
value

Fig. 6. Model performance of the CNN (a, c, d) and RNN (b, e, f) algorithms during the training phase.

management such that large-scale LSM is essential to the formation of 4.1. Models performance
landslide mitigation strategies. Since landslide frequency and intensity
are already high and are increasing in Iran, with many destructive To produce a reliable landslide susceptibility map, it is vital to
landslides in recent decades, a comprehensive and accurate delineate landslide-prone areas with high accuracy and to use a suitable
national-scale landslide susceptibility map of Iran is urgently required. metric for map evaluation. Low-accuracy hazard maps can lead to

513
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Fig. 7. Model performance of the CNN (a, c, d) and RNN (b, e, f) algorithms during the testing phase.

economic losses and catastrophic consequences. Appropriate models western parts of Iran are at higher risk of a landslide compare other
with high performance should be determined based on data availability regions.
and the characteristics and scale of the study area. The present study is
the first to apply state-of-the-art CNN and RNN deep learning algorithms,
4.2. Controlling landslides through geo-environmental management
which are appropriate for big datasets, to create a national-scale land-
slide susceptibility map of Iran. Dehnavi et al. (2015) applied an ANFIS
Except models prediction power, data quality and input variable se-
algorithm combined with an empirical model to map landslide suscep-
lection have significant effects on model results. We used the IGRF
tibility in Iran and reported 80% model accuracy, whereas the algorithms
technique to select the effective conditioning factors, as well as remove
used in our study showed accuracy values of 85% and 88% for the CNN
those with the null effect. Based on obtained results in this study, all
and RNN algorithms, respectively. It shows RNN and CNN algorithms
considered landslide conditioning factors are effective and the CNN- and
respectively have a 9.5% and 6% higher performance than then ANFIS
RNN-based models provided a comprehensive and accurate view of
algorithm. The ANFIS algorithm has difficulty in correctly determining
landslide hazard in different parts of Iran. Generally, geo-hazards like
membership function weights and is not recommended for national-scale
landslides can be more easily managed if decision-makers investigate and
applications. Both models showed that northern, northeastern, and
recognize the most effective geo-environment factors on landslide

514
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Fig. 8. Landslide susceptibility maps produced using the (a) CNN and (b) RNN algorithms.

occurrences. Our visual analysis revealed a link between landslide unit covered most areas in west and north of Iran and because of low
occurrence and conditioning factors, such that very high landslide sus- hardness (i.e. siltstone, sandstone, limestone and etc.) as well as having
ceptibility is associated with slope, altitude, and rainfall levels exceeding strata structures, they are very prone to landslide occurrences. Land use
20%, 2600 m, and 800 mm, respectively. Also this is in accordance with and distance from fault recognized as the third and fourth most effective
the IGRF technique, which slope degree is the most effective parameter factors with high degree of importance. Land use is the factor that affects
on landslide occurrences. The second highest parameter which effects on slope stabilization through the root system. Faults are defined as tectonic
landslide occurrences is geology. This factor has a great effect on rock fractures, which cause rock strength to be decrease and increase slope
hardness and accelerates the weathering process. The analysis showed instability (Bucci et al., 2016). When the slope is high, they are complete
that the sedimentary unit covered areas with high susceptibility. This in accordance with the nature of landslide occurrences and the land is

515
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Fig. 9. Model validation for the CNN and RNN algorithms using the (a) training and (b) testing datasets.

Fig. 12. Percentage of cities in Iran for each landslide risk category.

Fig. 10. Percentage of land area in Iran for each landslide susceptibil-
ity category.

Fig. 11. Percentage areas of each province to landslide hazard.

516
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Fig. 13. Percentage of cities to different landslide exposure severity in each province.

loose alongside with fracture which is completely prone to landslide. It is vital to continue to apply new techniques to LSM because com-
Also, our result in some extent is in accordance with Pourghasemi and plex, nonlinear phenomena cause landslides. Landslide prediction
Rahmati (2018), which stated that landslide occurrences in Mazandaran models provide homogenized data that can be more easily interpreted for
Province, Iran, strongly depends on the slope aspect, land use, and ge- comprehensive management of natural disasters in different environ-
ology. It is better to bear in mind that this is not true in all cases and ments; however, we must continue to improve their accuracy. The
according to the catchment characteristic, the importance of these con- delineation of areas with hazardous environmental conditions in Iran is a
ditioning factors are changeable. For instance, Pham et al. (2019) stated necessary step to identifying critical zones where the exploitation of
that distance from the road is the most effective factor in landslide natural resources and settlement expansion should be avoided (Pour-
occurrence at Uttarakhand Area, India. Chen et al. (2019a) demonstrated ghasemi et al., 2019). Achieved maps alongside importance of condi-
that altitude, distance to road and fault are the three most effective fac- tioning factor give an insight to decision-makers to landslide hazard
tors on landslide occurrence at Shangnan County, Shaanxi Province, mitigation through the relationship between landslide occurrences and
China. geo-environmental factors. Responsible management of these zones can
decrease the risk of human and economic losses and achieve sustainable
4.3. Previous research, limitation and future recommendation occupancy in high-risk regions of Iran. One of the main limitations of the
current study is lack of some effective data including soil depth, soil
At smaller spatial scales, Ding et al. (2016) applied a CNN-based texture and distance from water table. It is recommended to use these
model to identify landslides efficiently by detecting texture changes, factors for future research in order to enhance prediction power of al-
with minimal error (Ding et al., 2016). Sameen et al. (2020) evaluated gorithms, have more accurate landslide susceptibility maps as well as
landslide susceptibility using various models. They reported that a find a relationship between landslide occurrence with these new type of
CNN-based model provided the highest accuracy for validation data geo-environmental factors. Also recommended to hybridized deep
(83.11%) and the highest AUC (0.880). However, Ghorbanzadeh et al. learning algorithms with metaheuristic algorithm to optimize models
(2019a) suggested that because CNN-based models are often based on parameter and improve prediction capability of algorithms.
predefined measurable factors and/or settings selected using trial and
error, their results should be interpreted with caution. The performance 5. Conclusion
of CNN and RNN algorithms mainly depends on their design, including
training strategy, input window size, and layer depth (Ghorbanzadeh and Two well-known deep learning algorithms namely CNN- and RNN-
Blaschke, 2019). These algorithms produce consistent output sequences based models applied to produce a national-scale landslide susceptibil-
(Mezaal et al., 2017), but have error sources that are generally unrec- ity map of Iran. Overall, the RNN algorithm showed 3.4% higher per-
ognized or unacknowledged. RNN algorithms have been shown to formance compare to CNN. The results indicated that slope degree is the
improve landslide prediction accuracy effectively (Chen et al., 2015); most effective variable on landslide occurrences in Iran followed by ge-
those with optimized hyper parameters provide high-quality classifica- ology, land-use, distance from fault, plan curvature, rainfall, distance
tion outputs and have the highest potential for use in LSM of tropical from road, altitude, profile curvature, distance from rive and aspect. High
areas (Mezaal et al., 2017). The high estimation capability of RNN for and very high landslide susceptibility was detected for 20% of the land
LSM was emphasized by Mutlu et al. (2019), who found that enhancing area of Iran, mainly in the northern, northwestern, and western regions.
RNN complexity allowed the associated model to predict landslide The largest percentages of high or very high landslide susceptibility
initiation points correctly. The RNN algorithm also has a long short-term zones were found in the provinces of Chaharmahal and Bakhtiari, Gilan,
memory, such that it remembers all previous inputs during model pro- Kohgiluyeh and Boyer-Ahmad, Lorestan, Kurdistan, Mazandaran, Ker-
cessing; the incorporation of convolutional layers can stretch its effi- manshah, West Azarbayjan, and Zanjan, and 31% of all Iranian cities are
ciency to nearby pixels. RNN algorithm benefits from parameter sharing located in high or very high landslide susceptibility regions. This research
technique and it means RNN shares the parameters across different time contributes to better understanding of natural hazards and as tested, it
steps. can be applied to other regions of the world.

517
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Declaration of competing interest Fang, W., Ding, Y., Zhang, F., Sheng, V.S., 2019. DOG: a new background removal for
object recognition from images. Neurocomputing 361, 85–91.
Feizizadeh, B., Jankowski, P., Blaschke, T., 2014. A GIS based spatially-explicit sensitivity
The authors declare that they have no known competing financial and uncertainty analysis approach for multi-criteria decision analysis. Comput.
interests or personal relationships that could have appeared to influence Geosci. 64, 81–95.

Felicísimo, A.M., Cuartero, A., Remondo, J., Quir os, E., 2013. Mapping landslide
the work reported in this paper.
susceptibility with logistic regression, multiple adaptive regression splines,
classification and regression trees, and maximum entropy methods: a comparative
Acknowledgment study. Landslides 10 (2), 175–189.
Ghorbanzadeh, O., Blaschke, T., 2019. Optimizing sample patches selection of cnn to
improve the miou on landslide detection. In: Proceedings of the 5th International
This research was supported by the Basic Research Project of the Conference on Geographical Information Systems Theory, Applications and
Korea Institute of Geoscience and Mineral Resources (KIGAM) and Management: GISTAM, Heraklion, Greece, pp. 3–5.
Project of Environmental Business Big Data Platform and Center Con- Ghorbanzadeh, O., Blaschke, T., Aryal, J., Gholaminia, K., 2018a. A new GIS-based
technique using an adaptive neuro-fuzzy inference system for land subsidence
struction funded by the Ministry of Science and ICT. susceptibility mapping. Spatial Sci. https://doi.org/10.1080/
14498596.2018.1505564.
References Ghorbanzadeh, O., Blaschke, T., Gholamnia, K., Meena, S.R., Tiede, D., Aryal, J., 2019a.
Evaluation of different machine learning methods and deep-learning convolutional
neural networks for landslide detection. Rem. Sens. 11 (2), 196.
Akgun, A., Sezer, E.A., Nefeslioglu, H.A., Gokceoglu, C., Pradhan, B., 2012. An easy-to-use
Ghorbanzadeh, O., Feizizadeh, B., Blaschke, T., 2018b. An interval matrix method used to
MATLAB program (MamLand) for the assessment of landslide susceptibility using a
optimize the decision matrix in AHP technique for land subsidence susceptibility
Mamdani fuzzy algorithm. Comput. Geosci. 38 (1), 23–34.
mapping. Environmental Earth Sciences 77 (16), 584.
Alilou, H., Rahmati, O., Singh, V.P., Choubin, B., Pradhan, B., Keesstra, S., Ghiasi, S.S.,
Ghorbanzadeh, O., Meena, S.R., Blaschke, T., Aryal, J., 2019b. UAV-based slope failure
Sadeghi, S.H., 2019. Evaluation of watershed health using Fuzzy-ANP approach
detection using deep-learning convolutional neural networks. Rem. Sens. 11 (17),
considering geo-environmental and topo-hydrological criteria. J. Environ. Manag.
2046.
232, 22–36.
Guirado, E., Tabik, S., Alcaraz-Segura, D., Cabello, J., Herrera, F., 2017. Deep-learning
Bengio, Y., LeCun, Y., Henderson, D., 1994. Globally trained handwritten word recognizer
Convolutional Neural Networks for Scattered Shrub Detection with Google Earth
using spatial representation, convolutional neural networks, and hidden Markov
Imagery. Remote Sensing 9(12), 1220.
models. In: Advances in neural information processing systems, 937–944.
Guzzetti, F., Mondini, A.C., Cardinali, M., Fiorucci, F., Santangelo, M., Chang, K.T., 2012.
Bucci, F., Santangelo, M., Cardinali, M., Fiorucci, F., Guzzetti, F., 2016. Landslide
Landslide inventory maps: new tools for an old problem. Earth Sci. Rev. 112 (1–2),
distribution and size in response to Quaternary fault activity: the Peloritani Range,
42–66.
NE Sicily, Italy. Earth Surf. Process. Landforms 41 (5), 711–720.
 Jaeger, H., 2002. Tutorial on Training Recurrent Neural Networks, Covering BPPT, RTRL,
Bui, D.T., Hoang, N.D., Martínez-Alvarez, F., Ngo, P.T.T., Hoa, P.V., Pham, T.D.,
EKF and the“ Echo State Network” Approach. GMD Report 159. German National
Samui, P., Costache, R., 2020a. A novel deep learning neural network approach for
Research Center for Information Technology, p. 48.
predicting flash flood susceptibility: a case study at a high frequency tropical storm
Keesstra, S.D., Bouma, J., Wallinga, J., Tittonell, P., Smith, P., Cerda, A., Montanarella, L.,
area. Sci. Total Environ. 701, 134413.
Quinton, J.N., Pachepsky, Y., Van Der Putten, W.H., Bardgett, R.D., 2016. The
Bui, D.T., Pradhan, B., Lofman, O., Revhaug, I., Dick, O.B., 2012. Spatial prediction of
significance of soils and soil science towards realization of the United Nations
landslide hazards in Hoa Binh province (Vietnam): a comparative assessment of the
Sustainable Development Goals. SOIL 2(2), 111–128. https://doi.org/10.5194/soil-2
efficacy of evidential belief functions and fuzzy logic models. Catena 96, 28–40.
-111-2016.
Bui, D.T., Shirzadi, A., Shahabi, H., Chapi, K., Omidavr, E., Pham, B.T., Talebpour Asl, D.,
Khosravi, K., Sartaj, M., Tsai, F.T.C., Singh, V.P., Kazakis, N., Melesse, A.M., Prakash, I.,
Khaledian, H., Pradhan, B., Panahi, M., Bin Ahmad, B., 2019. A novel ensemble
Bui, D.T., Pham, B.T., 2018. A comparison study of DRASTIC methods with various
artificial intelligence approach for gully erosion mapping in a semi-arid watershed
objective methods for groundwater vulnerability assessment. Sci. Total Environ. 642,
(Iran). Sensors 19 (11), 2444.
1032–1049.
Bui, D.T., Tsangaratos, P., Nguyen, V.-T., Van Liem, N., Trinh, P.T., 2020b. Comparing the
Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. Imagenet classification with deep
prediction performance of a Deep Learning Neural Network model with conventional
convolutional neural networks. Adv. Neural Inf. Process. Syst. 1097–1105. https://
machine learning models in landslide susceptibility assessment. Catena 188, 104426.
doi.org/10.1145/3065386.
Can, R., Kocaman, S., Gokceoglu, C., 2019. A convolutional neural network architecture
LeCun, Y., Bengio, Y., 1995. Convolutional networks for images, speech, and time series.
for auto-detection of landslide photographs to assess citizen science and volunteered
The handbook of brain theory and neural networks 3361 (10), 1995.
geographic information data quality. ISPRS Int. J. Geo-Inf. 8 (7), 300.
LeCun, Y., Bottou, L., Bengio, Y., Haffner, P., 1998. Gradient-based learning applied to
Chapi, K., Singh, V.P., Shirzadi, A., Shahabi, H., Bui, D.T., Pham, B.T., Khosravi, K., 2017.
document recognition. Proc. IEEE 86 (11), 2278–2324.
A novel hybrid artificial intelligence approach for flood susceptibility assessment.
Lee, J.G., Jun, S., Cho, Y.W., Lee, H., Kim, G.B., Seo, J.B., Kim, N., 2017. Deep learning in
Environ. Model. Software 95, 229–245.
medical imaging: general overview. Korean J. Radiol. 18 (4), 570–584.
Chen, H., Zeng, Z., Tang, H., 2015. Landslide deformation prediction based on recurrent
Lee, S., Ryu, J.-H., Won, J.-S., Park, H.-J., 2004. Determination and application of the
neural network. Neural Process. Lett. 41 (2), 169–178.
weights for landslide susceptibility mapping using an artificial neural network. Eng.
Chen, W., Hong, H., Panahi, M., Shahabi, H., Wang, Y., Shirzadi, A., Pirasteh, S.,
Geol. 71 (3–4), 289–302.
Alesheikh, A.A., Khosravi, K., Panahi, S., Rezaie, F., 2019a. Spatial prediction of
Lima, P., Steger, S., Glade, T., Tilch, N., Schwarz, L., Kociu, A., 2017. Landslide
landslide susceptibility using gis-based data mining techniques of anfis with whale
Susceptibility Mapping at National Scale: a First Attempt for Austria, Workshop on
optimization algorithm (woa) and grey wolf optimizer (gwo). Appl. Sci. 9 (18), 3755.
World Landslide Forum. Springer, pp. 943–951.
Chen, W., Pourghasemi, H.R., Panahi, M., Kornejady, A., Wang, J., Xie, X., Cao, S., 2017.
Long, J., Shelhamer, E., Darrell, T., 2015. Fully convolutional networks for semantic
Spatial prediction of landslide susceptibility using an adaptive neuro-fuzzy inference
segmentation. In: Proceedings of the IEEE Conference on Computer Vision and
system combined with frequency ratio, generalized additive model, and support
Pattern Recognition, pp. 3431–3440.
vector machine techniques. Geomorphology 297, 69–85.
Mezaal, M.R., Pradhan, B., Sameen, M.I., Mohd Shafri, H.Z., Yusoff, Z.M., 2017.
Chen, W., Shahabi, H., Zhang, S., Khosravi, K., Shirzadi, A., Chapi, K., Pham, B.T.,
Optimized neural architecture for automatic landslide detection from high-resolution
Zhang, T., Zhang, L., Chai, H., Ma, J., 2018. Landslide susceptibility modeling based
airborne laser scanning data. Appl. Sci. 7 (7), 730.
on gis and novel bagging-based kernel logistic regression. Appl. Sci. 8 (12), 2540.
Mutlu, B., Nefeslioglu, H.A., Sezer, E.A., Akcayol, M.A., Gokceoglu, C., 2019. An
Chen, W., Zhao, X., Shahabi, H., Shirzadi, A., Khosravi, K., Chai, H., Zhang, S., Zhang, L.,
experimental research on the use of recurrent neural networks in landslide
Ma, J., Chen, Y., Wang, X., 2019b. Spatial prediction of landslide susceptibility by
susceptibility mapping. ISPRS Int. J. Geo-Inf. 8 (12), 578.
combining evidential belief function, logistic regression and logistic model tree.
Nachappa, T.G., Piralilou, S.T., Ghorbanzadeh, O., Shahabi, H., Blaschke, T., 2019.
Geocarto Int. 34 (11), 1177–1201.
Landslide susceptibility mapping for Austria using geons and optimization with the
Confuorto, P., Di Martire, D., Infante, D., Novellino, A., Papa, R., Calcaterra, D.,
dempster-shafer theory. Appl. Sci. 9 (24), 5393.
Ramondini, M., 2019. Monitoring of remedial works performance on landslide-
Nair, V., Hinton, G.E., 2010. Rectified linear units improve restricted Boltzmann
affected areas through ground- and satellite-based techniques. Catena 178, 77–89.
machines. In: Proceedings of the 27th International Conference on Machine Learning.
Dareshoori, F., Kasraian, N., 1998. Nature of Iran. Rouzaneh Kar Publication, Tehran,
ICML-10, pp. 807–814.
p. 198p (in Persian).
Nampak, H., Pradhan, B., Manap, M.A., 2014. Application of GIS based data driven
Dehnavi, A., Aghdam, I.N., Pradhan, B., Varzandeh, M.H.M., 2015. A new hybrid model
evidential belief function model to predict groundwater potential zonation. J. Hydrol.
using step-wise weight assessment ratio analysis (SWARA) technique and adaptive
513, 283–300.
neuro-fuzzy inference system (ANFIS) for regional landslide hazard assessment in
Neaupane, K.M., Piantanakulchai, M., 2006. Analytic network process model for landslide
Iran. Catena 135, 122–148.
hazard zonation. Eng. Geol. 85 (3–4), 281–294.
Ding, A., Zhang, Q., Zhou, X., Dai, B., 2016. Automatic recognition of landslide based on
Pham, B.T., Prakash, I., Khosravi, K., Chapi, K., Trinh, P.T., Ngo, T.Q., Hosseini, S.V.,
CNN and texture change detection. In: Proceedings of the 2016 31st Youth Academic
Bui, D.T., 2019. A comparison of Support Vector Machines and Bayesian algorithms
Annual Conference of Chinese Association of Automation (YAC). IEEE, Wuhan,
for landslide susceptibility modelling. Geocarto Int. 34 (13), 1385–1407.
China, 11-13 November 2016, pp. 444–448.
Piralilou, S.T., Shahabi, H., Jarihani, B., Ghorbanzadeh, O., Blaschke, T., Gholamnia, K.,
Du, Y., Wang, W., Wang, L., 2015. Hierarchical recurrent neural network for skeleton
Meena, S.R., Aryal, J., 2019. Landslide detection using multi-scale image
based action recognition. In: Proceedings of the IEEE Conference on Computer Vision
segmentation and different machine learning models in the higher Himalayas. Rem.
and Pattern Recognition (CVPR), Boston, MA, pp. 1110–1118. https://doi.org/
Sens. 11 (21), 2575.
10.1109/CVPR.2015.7298714.

518
P.T. Thi Ngo et al. Geoscience Frontiers 12 (2021) 505–519

Pourghasemi, H.R., Gayen, A., Panahi, M., Rezaie, F., Blaschke, T., 2019. Multi-hazard Susceptibility Mapping at National Scale: the Italian Case Study, Landslide Science
probability assessment and mapping in Iran. Sci. Total Environ. 692, 556–571. and Practice. Springer, pp. 287–295.
Pourghasemi, H.R., Rahmati, O., 2018. Prediction of the landslide susceptibility: which Van Westen, C.J., 2013. Remote sensing and GIS for natural hazards assessment and
algorithm, which precision? Catena 162, 177–192. disaster risk management. Treatise on Geomorphology 3, 259–298.
Rahmati, O., Panahi, M., Ghiasi, S.S., Deo, R.C., Tiefenbacher, J.P., Pradhan, B., Wang, Y., Fang, Z., Hong, H., Peng, L., 2020a. Flood susceptibility mapping using
Jahani, A., Goshtasb, H., Kornejady, A., Shahabi, H., Shirzadi, A., 2020. Hybridized convolutional neural network frameworks. J. Hydrol. 582, 124482.
neural fuzzy ensembles for dust source modeling and prediction. Atmos. Environ., Wang, Y., Fang, Z., Wang, M., Peng, L., Hong, H., 2020b. Comparative study of landslide
117320 susceptibility mapping with different recurrent neural networks. Comput. Geosci.,
Sameen, M.I., Pradhan, B., Lee, S., 2020. Application of convolutional neural networks 104445
featuring Bayesian optimization for landslide susceptibility assessment. Catena 186, Wang, Y., Hong, H., Chen, W., Li, S., Panahi, M., Khosravi, K., Shirzadi, A., Shahabi, H.,
104249. Panahi, S., Costache, R., 2019. Flood susceptibility mapping in Dingnan County
Sauter, T., Weitzenkamp, B., Schneider, C., 2010. Spatio-temporal prediction of snow (China) using adaptive neuro-fuzzy inference system with biogeography based
cover in the Black Forest mountain range using remote sensing and a recurrent neural optimization and imperialistic competitive algorithm. J. Environ. Manag. 247,
network. Int. J. Climatol. 30 (15), 2330–2341. 712–729.
Shirzadi, A., Solaimani, K., Roshan, M.H., Kavian, A., Chapi, K., Shahabi, H., Keesstra, S., Wilde, M., Günther, A., Reichenbach, P., Malet, J.-P., Hervas, J., 2018. Pan-European
Ahmad, B.B., Bui, D.T., 2019. Uncertainties of prediction accuracy in shallow landslide susceptibility mapping: ELSUS Version 2. J. Maps 14 (2), 97–104.
landslide modeling: sample size and raster resolution. Catena 178, 172–188. Yesilnacar, E.K., 2005. The Application of Computational Intelligence to Landslide
Siegelmann, H.T., Sontag, E.D., 1991. Turing computability with neural nets. Appl. Math. Susceptibility Mapping in Turkey. University of Melbourne, Department, p. 200.
Lett. 4 (6), 77–80. Yin, W., Kann, K., Yu, M., Schütze, H., 2017. Comparative Study of Cnn and Rnn for
Tehrany, M.S., Pradhan, B., Jebur, M.N., 2015. Flood susceptibility analysis and its Natural Language Processing arXiv preprint arXiv:1702.01923.
verification using a novel ensemble support vector machine and frequency ratio Zaremba, W., Sutskever, I., Vinyals, O., 2014. Recurrent Neural Network Regularization
method. Stoch. Environ. Res. Risk Assess. 29 (4), 1149–1165. arXiv preprint arXiv:1409.2329.
Trigila, A., Frattini, P., Casagli, N., Catani, F., Crosta, G., Esposito, C., Iadanza, C.,
Lagomarsino, D., Mugnozza, G.S., Segoni, S., Spizzichino, D., 2013. Landslide

519

You might also like