









































Pa
ge

 
1



Pa
ge

 
11

American Journal of  
Geospatial Technology (AJGT)

Landslide Prediction and Mapping through Geospatial and Neural Network Approach
Saurabh Kumar Anuragi1*, D. Kishan1

Volume 4 Issue 1, Year 2025
ISSN: 2833-8006 (Online)

DOI: https://doi.org/10.54536/ajgt.v4i1.4122
https://journals.e-palli.com/home/index.php/ajgt

Article Information ABSTRACT

Received: November 30, 2024

Accepted: January 04, 2025

Published: February 12, 2025

Identifying landslides and creating susceptibility maps are crucial in providing planners, local 
officials, and decision-makers with essential tools for effective disaster management strate-
gies. The accuracy of  these maps is vital in minimizing potential loss of  life and property. In 
order to develop comprehensive landslide susceptibility mappings, it is important to consider 
a range of  factors that encompass both terrain characteristics and meteorological conditions. 
Numerous advanced algorithms have been explored in the literature to enhance the preci-
sion of  these maps. This study utilizes a multi-layer perceptron neural network (MLPNN) 
with various activation functions, including ReLU, logistic, tanh, and identity, to compare 
model performance and establish the most accurate and reliable model for landslide suscep-
tibility mapping. Nine conditioning factors were analyzed, including aspect, elevation, land 
use/land cover, normalized difference vegetation index (NDVI), rainfall, slope, soil type, 
earthquake, and lithology. The performance of  the models was assessed using multiple met-
rics, including training score, testing score, kappa coefficient, specificity, sensitivity, and Area 
Under the Curve (AUC). The findings indicate that the MLPNN_logistic outperformed the 
other models, achieving kappa and AUC values of  0.504 and 0.757, respectively, in the de-
velopment of  susceptibility maps. As a result, the MLPNN_logistic model is identified as the 
most reliable and effective tool for landslide susceptibility mapping in this study, rendering it 
an optimal choice for predictive analyses in this field.

Keywords
Area Under Curve, Kappa, 
Landslide, Machine Learning, 
Multi Layer Perceptron

1 Department of  Civil Engineering, Maulana Azad National Institute of  Technology, Bhopal, India
* Corresponding author’s e-mail: saurabh.1399@gmail.com

INTRODUCTION
Landslides represent one of  the most catastrophic global 
geo-hazards, characterized by their complex geological 
nature and diverse occurrence across various geospatial 
environments and geomaterials. A landslide is defined as 
the mass movement of  soil, rock, or debris down a slope, 
which involves shear displacement along one or multiple 
slip surfaces. These slip surfaces may range from clearly 
visible, such as a well-defined sliding plane, to those 
that are inferred based on the geologic and geomorphic 
conditions surrounding the area. According to the United 
Nations Development Program (UNDP), landslides rank 
as the second largest natural disaster worldwide, leading 
to significant human casualties and extensive property 
damage (Azarafza et al., 2018; Pham et al., 2020). The 
devastating impact of  landslides underscores the urgent 
need to identify regions that are particularly prone to these 
events. Such identification is crucial for enhancing public 
safety and mitigating the adverse economic effects that 
can accompany landslide occurrences at both regional 
and national levels. In recent years, the study of  landslide 
susceptibility zones has gained prominence in the field of  
hazard management. This area of  research is dedicated 
to the development of  precise and up-to-date landslide 
susceptibility maps, which are essential tools for various 
stakeholders, including government agencies, urban 
planners, decision-makers, and local landowners. By 
providing detailed insights into areas at risk, these maps 
empower authorities to create comprehensive emergency 
response plans designed to reduce the detrimental 

impacts of  landslides on infrastructure, residential and 
commercial buildings, and the safety of  human life. 
Furthermore, mapping landslide-susceptible areas is not 
merely a preventive measure but a strategic approach to 
managing the potential repercussions of  landslides in 
vulnerable regions (Lee et al., 2004; Feizizadeh et al., 2014; 
Peethambaran et al., 2020; Murthy et al., 2023; Afroz et al., 
2022). The process of  assessing landslide susceptibility, 
however, is inherently complex. It typically entails a 
thorough investigation into numerous underlying factors 
that contribute to susceptibility—such as topography, 
geology, hydrology, land use, and climatic conditions—to 
produce zonation maps that delineate susceptible regions 
with a high degree of  spatial precision. Consequently, 
such rigorous assessments are critical for effective land 
use planning, risk reduction, and the formulation of  
policies aimed at safeguarding communities from the 
threats posed by landslides. 

LITERATURE REVIEW
Numerous studies have been conducted for landslide 
susceptibility mapping by various researchers utilizing a 
wide range of  statistical and machine-learning techniques 
documented in the literature. These approaches aim 
to identify areas prone to landslides by analyzing 
a combination of  geospatial, environmental, and 
geotechnical factors. Statistical techniques, such as logistic 
regression, frequency ratio, and weights of  evidence, 
have been extensively employed due to their simplicity 
and effectiveness in correlating landslide occurrences 



Pa
ge

 
12

https://journals.e-palli.com/home/index.php/ajgt

Am. J. Geo Spat. Technol. 4(1) 11-21, 2025

with causative factors. These models rely on historical 
landslide data to establish probabilistic relationships and 
are particularly useful for generating susceptibility maps 
in regions with abundant and reliable datasets. Khatun 
et al. (2022) employed a weighted overlay approach for 
landslide susceptibility mapping in Rangmati, Bangladesh, 
taking into account various conditioning factors such 
as soil texture, geology, lineament, slope, land use, and 
aspect. Arabameri et al. (2018) conducted a comparative 
study examining the analytic hierarchy process (AHP), 
frequency ratio (FR), index of  entropy (IOE), and weight 
of  evidence (WoE) methods for landslide risk zonation in 
Semnan Province, Iran, considering thirteen conditioning 
factors. In another study, Arabameri et al. (2018)  compared 
evidential belief  function (EBF), logistic regression (LR), 
and an ensemble of  LR-EBF for landslide zonation in 
the same region. Bopche et al. (2022) utilized the weight 
of  evidence method to create a landslide zonation 
map for Pune district, Maharashtra, considering ten 
conditioning factors. Tang et al. (2021) performed a 
comparative analysis between analytical hierarchy process 
information value (AHPIV) and logistic regression (LR) 
for landslide risk mapping in Zhushan County, China, 
incorporating thirteen conditioning factors. Kayastha et 
al. (2013) applied the analytical hierarchy process (AHP) 
for landslide zonation in the Tinau watershed, Nepal, 
focusing on geological, topographical, and hydrological 
factors. Regmi et al. (2014) compared the weight of  
evidence (WoE) and frequency ratio (FR) methods for 
landslide mapping in the Bhalubang-Shiwapur section 
of  the Mahendra Highway in Western Nepal. Mondal 
et al. (2014) integrated the frequency ratio (FR) with the 
analytical hierarchy model (AHP) to develop a hybrid 
model for landslide zonation in the Shiv-Khola watershed, 
Darjeeling Himalaya, based on ten conditioning factors. 
The advancement of  machine learning techniques, 
including support vector machines (SVM), random forest 
(RF), artificial neural networks (ANN), and gradient 
boosting machines (GBM), has gathered significant 
attention in recent years due to their capacity to effectively 
manage complex, nonlinear relationships between input 
variables and landslide occurrences. These methodologies 
frequently demonstrate superior performance compared 
to traditional statistical models, particularly when applied 
to large and diverse datasets. Additionally, the emergence 
of  hybrid models that combine statistical and machine 
learning approaches has proven beneficial, as they leverage 
the strengths of  both paradigms. Kavzoglu et al. (2013) 
compared support vector machine (SVM) approaches, 
criteria decision analysis, and logistic regression for 
landslide susceptibility mapping in Trabzon Province, 
Turkey. Kavzoglu et al. (2022) conducted a comparative 
study examining the efficacy of  extreme gradient 
boosting (XGBoost), random forest (RF), and natural 
gradient boosting (NGBoost) for landslide zonation 
mapping in Trabzon Province, Turkey. Park et al. (2012) 
performed a comparative analysis involving logistic 
regression, frequency ratio, artificial neural networks, and 

the analytical hierarchy process for landslide zonation in 
the Inje area of  Korea. Additionally, Kavzoglu et al. (2015) 
investigated various methodologies including logistic 
regression (LR), weight of  evidence (WoE), support 
vector regression (SVR), decision trees (DT), frequency 
ratio (FR), and statistical index for landslide susceptibility 
mapping in the Duzkoy district of  Turkey. Colkesen et al. 
(2016) conducted a similar comparative study assessing 
the effectiveness of  logistic regression, kernel-based 
Gaussian processes, and support vector machines for 
landslide susceptibility mapping in the Tonya district of  
Turkey. Pham et al. (2016) explored multiple methods, 
including rotation forest, random forest, bagging, naïve 
Bayes, AdaBoost, and MultiBoost, for landslide mapping 
in the Luc Yen district of  Vietnam. Kalantar et al. (2017)  
compared logistic regression, support vector machines, 
and artificial neural networks for landslide zonation in 
Mazandaran Province, Iran. Park et al. (2019) conducted 
a comparative analysis of  boosted regression trees and 
random forests for landslide zonation in the Woomyeon 
Mountain region of  South Korea. Hong et al. (2020) 
examined ensemble methods, specifically Forest by 
Penalizing Attributes (FPA) in conjunction with Bagging 
and LogitBoost alternating decision trees (LADT) with 
Bagging, for landslide zonation in the Youfanggou 
district of  China. Furthermore, support vector machines 
were utilized to analyze the predictive capability of  
conditioning factors within this context. Karakas et al. 
(2022) conducted a comparative study between multi-
layer perceptron (MLP) and random forest methods for 
landslide zonation in Elazig, Turkey.
Recent studies in the literature predominantly focus on 
mitigating landslide risk by identifying zones susceptible 
to such events, primarily through the comparison 
of  various statistical and machine learning models. 
These studies underscore the significance of  selecting 
appropriate algorithms and feature sets to generate 
accurate susceptibility maps. Furthermore, a considerable 
number of  researchers have examined the variation 
of  hyperparameters in both standalone and ensemble 
models. However, a critical gap exists in the existing 
literature regarding the impact of  hyperparameter 
variations on the performance of  neural networks. 
This area has been inadequately addressed, leading to 
the potential for suboptimal models that may either 
overfit the training data or fail to effectively capture 
the underlying patterns within the dataset. Additionally, 
the absence of  comprehensive analysis concerning 
hyperparameter settings complicates the assessment of  
different models’ true capabilities or their suitability for 
specific datasets and geographical contexts. In response 
to this gap, the present study utilizes a multi-layer 
perceptron neural network (MLPNN) to investigate how 
variations in activation functions—namely ‘relu’, ‘logistic’, 
‘identity’, and ‘tanh’—influence model performance. This 
examination of  hyperparameter variations aims to provide 
valuable insights into the robustness and stability of  the 
models. Such analyses are essential for the development 



Pa
ge

 
13

https://journals.e-palli.com/home/index.php/ajgt

Am. J. Geo Spat. Technol. 4(1) 11-21, 2025

of  reliable models that can be utilized effectively in real-
world environments, where data availability and quality 
may vary significantly. The performance evaluation will 
be conducted utilizing ROC and kappa metrics, which 
will facilitate objective and comprehensive comparisons 
among the methods employed. This approach is expected 
to enhance the understanding of  model behavior and 
contribute to the advancement of  effective solutions in 
landslide risk management.

MATERIALS AND METHODS
Study Area
Sikkim, a small yet vital state nestled in the North-Eastern 
Himalayas of  India, spans 7,096 square kilometers 
of  striking and diverse landscapes. Featuring a young 
mountain system, the region boasts an array of  geological 
wonders while also presenting challenges due to its 

susceptibility to landslides and seismic activity (Figure 
1). The vibrant capital, Gangtok, serves as a central hub 
for exploring a remarkable range of  elevations from 300 
to 8,000 meters above sea level, with 66% of  the state’s 
terrain being mountainous and often capped with snow. 
The climate of  Sikkim is uniquely varied, transitioning 
from tropical to alpine, which fosters an extraordinary 
ecosystem. Notably, rainfall in Gangtok is substantial at 
3,494 mm, in stark contrast to the minimal 82 mm in 
Thangu. This climatic diversity supports a rich tapestry 
of  life, contributing to Sikkim’s multi-ethnic population 
and vibrant cultural heritage. As of  2011, the state 
represented less than 0.05% of  India’s total population, 
with a population density of  86 people per square 
kilometer. This blend of  diversity and stunning natural 
beauty makes Sikkim an important area for ecological and 
cultural exploration.

Figure 1: Study Area Map

However, the uneven population distribution in Gangtok 
has created a unique set of  challenges for urban 
development and service delivery. The rapid growth of  
the population in a confined area has led to increased 
pressure on transportation systems, housing availability, 
and modern commercial development. As a result, 
addressing traffic congestion, housing shortages, and 
the emergence of  informal settlements is a priority. It 
is essential for Gangtok to find sustainable solutions to 
balance development with environmental preservation. 
Geologically, the Sikkim Himalaya, particularly along 
the Teesta Valley, is defined by the Main Central Thrust 
(MCT) and Main Boundary Thrust (MBT) zones, which 
delineate various grades of  Himalayan rocks. The central 
crystalline area in North Sikkim features high-grade 
gneisses, migmatites, and granitic intrusions, highlighting 
the region’s complex geological history. Additionally, 
Sikkim’s position in Zone IV on India’s Seismic Zoning 

Map underscores the importance of  earthquake 
preparedness and sustainable planning to ensure the 
safety and resilience of  its communities.

Landslide Conditioning Factors (LCFs)
The study identified nine key conditioning factors 
essential for accurate landslide susceptibility mapping, 
along with an analysis of  their spatial distribution (Figure 
2). The slope map of  the study area has been categorized 
into five distinct classes: 0-15° (very low), 15-25° (low), 
25-35° (moderate), 35-45° (high), and 45-90° (very high). 
The aspect map has been classified into nine categories, 
including flat, north, northeast, east, southeast, south, 
southwest, west, and northwest. The elevation map 
is divided into three classes: (i) 222 – 2460 m, (ii) 2460 
– 4227 m, and (iii) 4227 – 7899 m. The land use/land 
cover (LULC) map is classified into nine categories: (i) 
Water, (ii) Trees, (iii) Flooded vegetation, (iv) Crops, (v) 



Pa
ge

 
14

https://journals.e-palli.com/home/index.php/ajgt

Am. J. Geo Spat. Technol. 4(1) 11-21, 2025

Built area, (vi) Bare ground, (vii) Snow/ice, (viii) Clouds, 
and (ix) Rangeland. The normalized difference vegetation 
index (NDVI) map is classified into five classes: (i) (-) 
0.603 – 0.025, (ii) 0.026 – 0.125, (iii) 0.126 – 0.232, (iv) 
0.233 – 0.465, and (v) 0.466 – 1. The rainfall map is 
divided into five categories: (i) 1000 - 1647 mm, (ii) 1648 
- 2235 mm, (iii) 2236 - 2729 mm, (iv) 2730 - 3353 mm, 
and (v) 3354 - 4000 mm. The soil type map has been 
classified into five categories: (i) Humid Acrisols, (ii) 
Dystric Cambisols, (iii) Gleysols luvi Soils, (iv) Lithosols, 
and (v) Dystric Regosols. The lithology map has been 
classified into seven categories (i) Jurassic metamorphic 
and sedimentary rocks, (ii) Cretaceous and undivided 
igneous rocks, (iii) Mesozoic and Paleozoic intrusive and 
metamorphic rocks, (iv) undivided precambrian rocks, 
(v) undivided paleozoic rocks, (vi) Quaternary sediments 
and (vii) Tertiary and cretaceous sedimentary rocks. 
Lastly, the earthquake map is categorized into three 
magnitude classes: (i) 0.011 - 3.291, (ii) 3.292 - 6.018, 
and (iii) 6.019 - 11.794. These thematic maps were 
precisely prepared to enhance the analysis of  landslide 
susceptibility by integrating them with landslide 
occurrence data, thereby enabling a more precise and 
comprehensive assessment.

Landslide Inventory
The landslide data utilized for this analysis has been 
obtained from Bhukosh, Geological Survey of  India, 
encompassing a total of  693 data points, as depicted 
in Figure 3. These data points were imported into 
ArcGIS, where polygons were generated to establish 
a comprehensive dataset for further examination. To 
ensure the dataset’s balance, an additional 695 non-
landslide data points were randomly generated within 
ArcGIS, and corresponding polygons were created. By 
integrating the Landslide Conditioning Factors (LCFs) 
with both landslide and non-landslide data, a consolidated 
dataset comprising 12165 data points was developed. 
This dataset was then divided in a 70:30 ratio, with 70% 
allocated for training purposes and 30% for testing, to 
facilitate a thorough analysis.
The distribution and variability of  each input variable 
related to landslides are presented in Figure 4. The 
figure demonstrates notable correlations among various 
features, highlighting significant relationships among 
them. The correlation coefficient of  0.61 between Land 
Use and Land Cover (LULC) and Elevation underscores 
how vegetation patterns and human activities vary 
with changes in altitude. Additionally, a moderate 

Figure 2: Landslide Conditioning Factors Map



Pa
ge

 
15

https://journals.e-palli.com/home/index.php/ajgt

Am. J. Geo Spat. Technol. 4(1) 11-21, 2025

correlation of  0.53 between the Normalized Difference 
Vegetation Index (NDVI) and Rainfall reveals the 
impact of  precipitation on vegetation health and density. 
Furthermore, a correlation of  0.58 between Rainfall and 
Soil Type indicates that soil characteristics are influenced 
by rainfall distribution, which affects factors such as 
soil moisture and composition. Conversely, several 
features exhibit minimal correlation, suggesting weaker 
interdependence. For example, the weak correlation 
observed between Elevation and Soil Type suggests that soil 
properties are largely independent of  altitude. Similarly, the 

correlation between Rainfall and Elevation is low, indicating 
that rainfall patterns do not strongly relate to elevation 
variations within the region. The relationship between 
NDVI and Elevation is also weak, indicating that vegetation 
density is not significantly affected by altitude in this study 
area. These insights into feature correlations are essential 
for comprehending the interplay of  factors contributing 
to landslide susceptibility. Identifying highly correlated 
features can help reduce redundancy within models, while 
incorporating weakly correlated features ensures the diversity 
and independence of  factors considered in the analysis.

Figure 3: Landslide Inventory

Figure 4: Correlation matrix of  conditioning factors



Pa
ge

 
16

https://journals.e-palli.com/home/index.php/ajgt

Am. J. Geo Spat. Technol. 4(1) 11-21, 2025

Methodology
The methodology utilized in this study is presented in 
Figure 5, which outlines the systematic approach adopted 
to achieve reliable landslide susceptibility mapping. 
The research employs a Multi-Layer Perceptron Neural 
Network (MLPNN) to generate landslide susceptibility 
maps (LSMs) across a defined range of  hyperparameters, 
detailed in Table 1. A significant focus of  the study is 
the comparative analysis of  various activation functions, 
specifically ‘relu’, ‘logistic’, ‘tanh’, and ‘identity’, and their 
effectiveness in producing LSMs. The choice of  activation 
function is critical, as it influences the network’s ability to 
learn complex, non-linear relationships inherent in the 
input data.
This is especially relevant in landslide susceptibility 
mapping, where the interactions among geological, 
hydrological, and environmental factors are often 
intricate and non-linear. By varying the activation 

functions and evaluating their impact on the model’s 
predictive accuracy, this research aims to identify the 
most appropriate function for this particular application. 
Notably, the ReLU activation function is recognized 
for its computational efficiency and its effectiveness in 
overcoming the vanishing gradient problem, making 
it a widely used option in deep learning. Conversely, 
the sigmoid and tanh functions are more suitable for 
situations requiring probabilistic interpretations, as 
they constrain outputs within a defined range. The 
comparative analysis conducted in this study provides 
valuable insights into how different activation functions 
influence the overall performance of  the MLPNN. 
This is assessed through key metrics, including kappa, 
sensitivity, specificity, and AUC. The findings contribute 
to an enhanced understanding of  the model’s capabilities 
and offer guidance for the refinement of  methodologies 
in landslide susceptibility mapping.

Table 1: MLPNN hyperparameter settings
Hyperparameter Values
solver adam
activation relu; logistic; tanh; identity
hidden layer [50,50,50,50,50]
max_iter 500
learning_rate adaptive
learning_rate 0.01

Furthermore, hyperparameters are essential in optimizing 
the model’s performance. By systematically adjusting 
these parameters, we aim to prevent underfitting and 
overfitting, thereby enhancing the model’s effectiveness 
with the training data. This careful calibration has enabled 
us to generate a series of  Landslide Susceptibility Maps 
(LSMs), which were subsequently compared to identify 
the most effective model. These maps were visually 
represented and validated against ground-truth data 
as well as historical landslide occurrences, providing 

a comprehensive assessment of  their reliability. The 
comparative analysis yields critical insights into how 
various activation functions can influence the model’s 
performance and predictive accuracy. The methodology 
adopted in this study exemplifies a meticulous 
and systematic approach to landslide susceptibility 
mapping. By harnessing the capabilities of  MLPNN 
and investigating a wide range of  hyperparameters and 
activation functions, this research establishes a robust 
framework for the development of  high-quality LSMs.

Figure 5: Methodology of  the study



Pa
ge

 
17

https://journals.e-palli.com/home/index.php/ajgt

Am. J. Geo Spat. Technol. 4(1) 11-21, 2025

RESULTS AND DISCUSSION
The study utilized a Multi-Layer Perceptron Neural 
Network (MLPNN) to conduct a comparative analysis of  
various activation functions, specifically ‘relu’, ‘logistic’, 
‘tanh’, and ‘identity’, in the context of  generating 
landslide susceptibility maps. These activation functions 
were selected for their distinct characteristics and 
prevalent application in neural network methodologies, 
particularly for addressing non-linear classification 
challenges. The primary objective of  the study was to 
identify the activation function that is most effective for 
landslide susceptibility mapping by examining its impact 
on the overall performance of  the model. Optimal 
hyperparameters for each configuration—including the 
number of  hidden layers, neurons per layer, learning rate, 
and maximum iterations—were rigorously determined 
through systematic experimentation, as detailed in Table 
1. Careful fine-tuning of  these hyperparameters was 
conducted to ensure that the MLPNN accurately captured 
the underlying patterns within the input data while 
preventing both overfitting and underfitting. To facilitate 
a fair and unbiased evaluation, the selected configurations 
were consistently applied across all activation functions. 
The model’s performance was assessed using a range of  
metrics, including training score, testing score, specificity, 
sensitivity, kappa, and area under the curve (AUC). This 
comparative analysis of  the resulting maps provided 
valuable insights into the influence of  different activation 
functions on the model’s accuracy, robustness, and overall 
capacity to effectively predict landslide-prone areas.
The ROC curves presented in Figure 6 detail the 
classification performance of  the various models, 

with the AUC values underscoring their capacity to 
differentiate between classes. The AUC values for the 
models are as follows: 0.757 for MLPNN_logistic, 
0.751 for MLPNN_relu, 0.750 for MLPNN_tanh, and 
0.736 for MLPNN_identity. These results highlight the 
differing effectiveness of  the activation functions in 
distinguishing between landslide-prone and stable areas. 
In addition to the AUC analysis, Figure 7 provides a 
comprehensive overview of  the models performance 
across multiple metrics, including training scores, testing 
scores, sensitivity, specificity, and kappa coefficient. The 
MLPNN_relu model exhibited a commendable balance 
between training and testing performance, achieving 
scores of  0.742 and 0.740, respectively. Its high sensitivity 
of  0.908 indicates an excellent capacity for identifying 
landslide-prone areas; however, its specificity of  0.594 
suggests a higher incidence of  false positives. Conversely, 
the MLPNN_logistic model attained the highest kappa 
value of  0.504, signifying superior agreement between 
predicted and actual classifications. With a training score 
of  0.760, a testing score of  0.748, a sensitivity of  0.883, 
and a specificity of  0.631, the logistic activation function 
emerged as the most robust overall. The MLPNN_tanh 
function demonstrated consistent performance, with 
training and testing scores of  0.745 and 0.741, respectively. 
Its sensitivity of  0.875 and specificity of  0.624 render it 
a competitive option. While MLPNN_identity achieved 
the highest specificity of  0.673, indicating fewer false 
positives, it recorded the lowest sensitivity at 0.798, which 
hampers its effectiveness in identifying landslide-prone 
areas. The kappa value of  0.466 further reflects lower 
consistency compared to the other activation functions.

Figure 6: ROC curves of  MLPNN models



Pa
ge

 
18

https://journals.e-palli.com/home/index.php/ajgt

Am. J. Geo Spat. Technol. 4(1) 11-21, 2025

Figure 7: Evaluation metrics

The evaluation metrics, along with the confusion matrices 
presented in Figure 8, offer a comprehensive analysis of  
the predictive accuracy of  the models. These matrices 
detail the distribution of  true positives, true negatives, 
false positives, and false negatives, thereby providing 
valuable insights into the classification performance of  
each model. The confusion matrix for MLPNN_identity 
indicates the highest number of  misclassifications, 
totaling 979 instances. This finding suggests a relatively 
lower predictive accuracy, which is further reflected 
in its AUC score of  0.736. Similarly, MLPNN_relu 
shows 946 misclassifications, indicating limitations in 
its predictive capabilities despite its competitive AUC 
of  0.751 and commendable sensitivity. In comparison, 
MLPNN_tanh recorded 944 misclassifications, 
demonstrating slightly better performance than the 
ReLU activation function, yet still not achieving 
optimal classification results. Notably, MLPNN_logistic 
exhibited the lowest number of  misclassifications at 
918, showcasing superior classification abilities relative 
to the other activation functions. This outcome aligns 
with its highest AUC value of  0.757, further affirming 
its effectiveness in differentiating between classes. The 
reduced misclassification rate observed in MLPNN_
logistic emphasizes its potential for practical applications, 
particularly in contexts where minimizing errors is crucial 
for reliable landslide susceptibility mapping.
The landslide susceptibility maps generated mentioned 
earlier are illustrated in Figure 9. These maps categorize 
the terrain into distinct susceptibility levels, serving 
as an essential tool for identifying high-risk areas and 
informing effective mitigation strategies. Figure 10 
further quantifies the spatial distribution by depicting the 
percentage of  the study area encompassed within each 

susceptibility classification—very low, low, moderate, 
high, and very high—across the various machine learning 
models. This analysis provides valuable insights into how 
different activation functions influence the classification 
of  susceptibility zones. For the MLPNN_relu model, a 
significant portion of  the study area, 26.79%, is predicted 
to fall into the “very high” susceptibility class, indicating 
notable areas of  risk. In contrast, only 5.74% of  the 
area is classified as “very low,” which suggests a limited 
extent of  minimal risk zones. This pattern illustrates 
the model’s tendency to identify areas with heightened 
susceptibility. In the case of  the MLPNN_logistic model, 
the “high” susceptibility class occupies the largest area 
at 25.96%, while the “very low” class covers the smallest 
area at 6.27%. This distribution indicates a more balanced 
classification approach compared to other models, with a 
focus on moderately high-risk zones. For the MLPNN_
tanh model, the “low” susceptibility class is predominant, 
encompassing 27.46% of  the area, whereas the “very 
low” class represents only 6.42%.
This outcome underscores the model’s inclination to 
categorize a larger proportion of  the terrain as having 
lower susceptibility, which may reflect a conservative 
classification strategy. Lastly, the MLPNN_identity model 
exhibits a distinctive distribution, with the “moderate” 
susceptibility class accounting for the largest share at 
46.81%, and the “very low” class covering just 3.07% of  
the area. This model appears to categorize a substantial 
portion of  the terrain into a mid-range risk category, 
offering a unique perspective on susceptibility zoning. 
The variations in area distribution among these models 
highlight the significant impact of  activation functions 
and model parameters on the development of  landslide 
susceptibility maps.



Pa
ge

 
19

https://journals.e-palli.com/home/index.php/ajgt

Am. J. Geo Spat. Technol. 4(1) 11-21, 2025

Figure 8: Confusion matrix of  MLPNN models

Figure 9: Generated LSM by MLPNN



Pa
ge

 
20

https://journals.e-palli.com/home/index.php/ajgt

Am. J. Geo Spat. Technol. 4(1) 11-21, 2025

Overall, this analysis highlights MLPNN_logistic and 
MLPNN_relu as the most effective activation functions 
for producing accurate landslide susceptibility maps. 
Although logistic activation slightly surpassed ReLU in 
the overall metrics, ReLU’s superior sensitivity makes it 
particularly advantageous for minimizing false negatives 
in identifying landslide-prone zones. 

CONCLUSION
The identification of  regions susceptible to landslides, 
along with the precise determination of  their locations 
based on specified susceptibility levels, is essential for 
effective planning initiatives. Numerous methodologies 
have been proposed in the existing literature for the 
development of  landslide susceptibility maps. This study 
assesses the efficacy of  the Multilayer Perceptron Neural 
Network (MLPNN) in constructing accurate and reliable 
models for generating landslide susceptibility maps 
specifically for Sikkim, India. The research utilizes nine 
conditioning factors, which include aspect, elevation, 
land use/land cover, normalized difference vegetation 
index (NDVI), rainfall, slope, soil type, earthquake, and 
lithology. A comprehensive set of  evaluation metrics has 
been employed to assess and compare the performance 
of  the models. These metrics include training score, 
testing score, kappa coefficient, specificity, sensitivity, 
and Area Under the Curve (AUC). Such metrics provide 
a robust framework for analyzing the predictive accuracy 
and reliability of  each model in classifying landslide 
susceptibility. The findings indicate that the MLPNN_
logistic model outperformed other models based on the 
evaluation metrics. It achieved a kappa coefficient of  
0.504, indicating a strong correlation between predicted 
and actual classifications while accounting for chance 
agreement. Additionally, the MLPNN_logistic recorded 
the highest AUC value of  0.757, demonstrating its 
exceptional capability to differentiate between classes and 
reliably predict landslide susceptibility. 
Moreover, analysis of  the confusion matrix revealed that 

the MLPNN_logistic model had the lowest number of  
misclassifications, totalling 918 incorrectly classified 
instances, thus establishing it as the most accurate 
among the evaluated models. These results underscore 
the MLPNN_logistic model’s capability to minimize 
prediction errors and its robustness in managing complex 
datasets with multiple conditioning factors. In conclusion, 
the combination of  a high kappa coefficient, superior 
AUC, and minimal misclassifications positions the 
MLPNN_logistic model as a reliable and effective tool 
for landslide susceptibility mapping, making it an optimal 
choice for predictive analyses in this domain.

REFERENCES
Afroz, T., Miah, M. G., Abdullah, H. M., Islam, M. R., & 

Rahman, M. M. (2022). Monitoring of  LULC Changes   
and Forest Loss Using Geospatial Technique: A Case 
Study from Northern Region of  Bangladesh. American 
Journal of  Geospatial Technology, 1(2), 1–9. https://doi.
org/10.54536/ajgt.v1i2.907

Arabameri, A., Pradhan, B., Rezaei, K., Yamani, M., 
Pourghasemi, H. R., & Lombardo, L. (2018). Spatial 
modelling of  gully erosion using evidential belief  
function, logistic regression, and a new ensemble 
of  evidential belief  function–logistic regression 
algorithm. Land Degradation & Development, 29(11), 
4035-4049.

Arabameri, A., Rezaei, K., Pourghasemi, H. R., Lee, 
S., & Yamani, M. (2018). GIS-based gully erosion 
susceptibility mapping: a comparison among three 
data-driven models and AHP knowledge-based 
technique. Environmental earth sciences, 77, 1-22.

Azarafza, M., Ghazifard, A., Akgün, H., & Asghari-
Kaljahi, E. (2018). Landslide susceptibility assessment 
of  South Pars Special Zone, southwest Iran. 
Environmental Earth Sciences, 77, 1-29.

Bopche, L., & Rege, P. P. (2022). Landslide Susceptibility 
Mapping: An Integrated Approach using Geographic 
Information Value, Remote Sensing, and Weight of  

Figure 10: Area Distribution of  study area by MLPNN



Pa
ge

 
21

https://journals.e-palli.com/home/index.php/ajgt

Am. J. Geo Spat. Technol. 4(1) 11-21, 2025

Evidence Method. Geotechnical and Geological Engineering, 
40(6), 2935–2947. https://doi.org/10.1007/s10706-
022-02070-4

Colkesen, I., Sahin, E. K., & Kavzoglu, T. (2016). 
Susceptibility mapping of  shallow landslides using 
kernel-based Gaussian process, support vector 
machines and logistic regression. Journal of  African 
Earth Sciences, 118, 53-64.

Feizizadeh, B., & Blaschke, T. (2014). An uncertainty 
and sensitivity analysis approach for GIS-based 
multicriteria landslide susceptibility mapping. 
International Journal of  Geographical Information Science, 
28(3), 610-638.

Hong, H., Liu, J., & Zhu, A. X. (2020). Modeling landslide 
susceptibility using LogitBoost alternating decision 
trees and forest by penalizing attributes with the 
bagging ensemble. Science of  the Total Environment, 718, 
137231.

Kalantar, B., Pradhan, B., Naghibi, S. A., Motevalli, A., 
& Mansor, S. (2018). Assessment of  the effects of  
training data selection on the landslide susceptibility 
mapping: a comparison between support vector 
machine (SVM), logistic regression (LR) and artificial 
neural networks (ANN). Geomatics, Natural Hazards 
and Risk, 9(1), 49-69.

Karakas, G., Kocaman, S., & Gokceoglu, C. (2022). 
Comprehensive performance assessment of  landslide 
susceptibility mapping with MLP and random forest: 
a case study after Elazig earthquake (24 Jan 2020, Mw 
6.8), Turkey. Environmental Earth Sciences, 81(5), 144.

Kavzoglu, T., & Teke, A. (2022). Predictive Performances 
of  ensemble machine learning algorithms in landslide 
susceptibility mapping using random forest, extreme 
gradient boosting (XGBoost) and natural gradient 
boosting (NGBoost). Arabian Journal for Science and 
Engineering, 47(6), 7367-7385.

Kavzoglu, T., Kutlug Sahin, E., & Colkesen, I. (2015). An 
assessment of  multivariate and bivariate approaches 
in landslide susceptibility mapping: a case study of  
Duzkoy district. Natural Hazards, 76, 471-496.

Kavzoglu, T., Sahin, E. K., & Colkesen, I. (2014). Landslide 
susceptibility mapping using GIS-based multi-criteria 
decision analysis, support vector machines, and logistic 
regression. Landslides, 11, 425-439.

Kayastha, P., Dhital, M. R., & De Smedt, F. (2013). 
Application of  the analytical hierarchy process (AHP) 
for landslide susceptibility mapping: A case study 
from the Tinau watershed, west Nepal. Computers & 
Geosciences, 52, 398-408.

Khatun, M., Hossain, A. S., Sayem, H. M., Moniruzzaman, 
M., Ahmed, Z., & Rahaman, K. R. (2023). Landslide 
susceptibility mapping using weighted-overlay 
approach in Rangamati, Bangladesh. Earth Systems and 

Environment, 7(1), 223-235.
Lee, S., & Choi, J. (2004). Landslide susceptibility 

mapping using GIS and the weight-of-evidence 
model. International Journal of  Geographical Information 
Science, 18(8), 789-814.

Mondal, S., & Maiti, R. (2013). Integrating the analytical 
hierarchy process (AHP) and the frequency ratio (FR) 
model in landslide susceptibility mapping of  Shiv-
khola watershed, Darjeeling Himalaya. International 
Journal of  Disaster Risk Science, 4, 200-212.

Park, S., & Kim, J. (2019). Landslide susceptibility mapping 
based on random forest and boosted regression tree 
models, and a comparison of  their performance. 
Applied Sciences, 9(5), 942.

Park, S., Choi, C., Kim, B., & Kim, J. (2013). Landslide 
susceptibility mapping using frequency ratio, analytic 
hierarchy process, logistic regression, and artificial 
neural network methods at the Inje area, Korea. 
Environmental Earth Sciences, 68(5), 1443-1464.

Peethambaran, B., Anbalagan, R., Kanungo, D. P., 
Goswami, A., & Shihabudheen, K. V. (2020). A 
comparative evaluation of  supervised machine 
learning algorithms for township level landslide 
susceptibility zonation in parts of  Indian Himalayas. 
Catena, 195, 104751.

Pham, B. T., Bui, D. T., Dholakia, M. B., Prakash, I., 
Pham, H. V., Mehmood, K., & Le, H. Q. (2017). A 
novel ensemble classifier of  rotation forest and Naïve 
Bayer for landslide susceptibility assessment at the 
Luc Yen district, Yen Bai Province (Viet Nam) using 
GIS. Geomatics, Natural Hazards and Risk, 8(2), 649-
671.

Pham, V. D., Nguyen, Q. H., Nguyen, H. D., Pham, V. M., 
& Bui, Q. T. (2020). Convolutional neural network—
optimized moth flame algorithm for shallow landslide 
susceptible analysis. IEEE Access, 8, 32727-32736.

Ramana Murty, M. V., Ravi Kumar, C., Srinivasu, K., 
Kannan, R., & Sundar, B. (2023). Monitoring of  
Coastal Geo-Environment for Hazard Mitigation: 
A Case Study of  Machilipatnam Region, Andhra 
Pradesh, India. American Journal of  Geospatial Technology, 
1(2), 27–38. https://doi.org/10.54536/ajgt.v1i2.1381

Regmi, A. D., Yoshida, K., Pourghasemi, H. R., DhitaL, 
M. R., & Pradhan, B. (2014). Landslide susceptibility 
mapping along Bhalubang—Shiwapur area of  mid-
Western Nepal using frequency ratio and conditional 
probability models. Journal of  Mountain Science, 11, 
1266-1285.

Tang, R. X., Yan, E. C., Wen, T., Yin, X. M., & Tang, 
W. (2021). Comparison of  logistic regression, 
information value, and comprehensive evaluating 
model for landslide susceptibility mapping. 
Sustainability, 13(7), 3803.


