1The northern region of Pakistan is categorized into three prominent mountain ranges: the Hindukush, Karakoram, and Himalaya. Among these ranges, the Karakoram region faces frequent and significant landslides, representing a repeated natural disaster (Abbas et al., 2023). The increasing severity of landslide hazards in northern Pakistan poses a serious concern, attributable to several factors such as regional tectonic activity, topography, and the impact of climate change (Su et al., 2022). Furthermore, a large devastating landslide happened at Ataabad village in 2009, this landslide killed 19 human lives, displaced more than 6000 people, blocked the main Hunza River, and caused a 14 km natural lake. This landslide blocked the KKH, the main strategic route linking China and Pakistan (Kulsoom et al., 2023).
2Landslide susceptibility maps (LSMs) are essential to estimate landslide susceptibility in an area of various geological and climatic conditions (Alsabhan et al., 2022). Regardless of the plethora of methods and techniques that have been proposed and examined, there is not yet a uniform procedure for generating landslide susceptibility maps (Dikshit et al., 2020). Moreover, these landslide susceptibility methods can be separated into three classes: statistical, deterministic, and heuristic (Alsabhan et al., 2022). Moreover, heuristic methods have drawbacks, such as complexity leading to improper generalization and duplicability issues when assigning variable weights (Reichenbach et al., 2018). Recently, quantitative methods have become more prevalent due to their mathematical expressions and capability to identify the relationship between landslides and their causative factors (Bai et al., 2010). These statistical methods have been commonly used for LSM (Li et Chen, 2019) such as Frequency ratio (FR), Information value method (Khan et al., 2022), Weight of evidence (WofE) (Alsabhan et al., 2022), evidential belief function (Li et Chen, 2019), and regressive multiple analysis (Mandal et Mondal, 2019). These methods have established effectiveness in accurately assessing the likelihood of landslides occurring in specific areas (Qazi et al., 2023).
3Furthermore, the Weight of Evidence (WofE) model is a prominent bi-variate statistical approach for evaluating landslide susceptibility across various landslide-prone regions (Gupta et al., 2023). Additionally, the combination of (FR) and (WofE) offers few advantages in analyzing landslide susceptibility (LSM) (Khan et al., 2022). Additionally, these methods provide inclusive insights into the key factors responsible for landslide occurrence (Mersha et Meten, 2020). The integration of frequency ratios (FR) and weights of evidence (WofE) offers various advantages in the assessment of landslide susceptibility mapping (Qazi et al., 2023). However, it is worth noting that these methods often lack a validation step and do not consider data uncertainties such as measurement errors, this can introduce bias in the model outputs and potentially affect the accuracy of the resulting landslide susceptibility maps (Khan et al., 2022). To efficiently address the inherent lacunas of widely adopted models such as WofE and FR, it is imperative to explore and implement alternative approaches that can overcome the inadequacies associated with these models.
4Furthermore, according to Li et Chen (2020), machine learning methods (ML) have been executed by various researchers for LSM based on intelligence-based methods for landslide susceptibility zonation. Generally, used machine learning models include support vector machines (SVM) (Khan et al., 2022), artificial neural networks (Abbas et al., 2023), Naive Bayes and logistic model trees (Pham et al., 2016), decision tree (Pradhan, 2013), rotation forests (Ada et San, 2018), and random forest (RF) (Kulsoom et al., 2023). Pham et al. (2016) made a comparative analysis of the prediction accuracies of SVM, and Naive Bayes found SVM had the highest prediction accuracy. Ada and San compared SVM and RF for LSM and stated both were capable of LSM (Ada et San, 2018). Merghadi et al. (2020), found that, compared to other machine learning models, Tree-based RF achieved superior performance in LSM. However, despite these advancements in DMLTs and statistical-based modeling, there is still a lack of consensus on the best method for LSM, However, the effectiveness of the models merely depends upon the geo-climatic conditions of the pertinent study area (Abbas et al., 2023). Furthermore, the pertinency of these techniques and methods in diverse geological and climatic settings remains ambiguous and challenging (Alsabhan et al., 2022). Hence, it is of vital importance to gage the dependability of landslide susceptibility methods and techniques to ascertain the precision, viability, and appropriateness of the models employed in different geological and climatic scenarios (Khan et al., 2022).
5The primary aim of this study is to assess and precisely evaluate the spatial correlation between factors influencing landslide occurrence. To accomplish this, a model called Intuitionistic Fuzzy Divergence is proposed, which utilizes Fuzzy systems. To ensure accurate, precise, and valid results, two conventional methods, (FR) and (WofE), are also employed for comparison purposes. The main objective of this research is to determine the most suitable and effective technique for landslide susceptibility modeling in the study area. To achieve this, three advanced DMLTs (DeepLabV3+, InceptionV3+, and U-Net) are compared with three statistical models (WofE, FR, and IF-D). This study also believes that the findings of this work make a significant contribution to the scientific community, and the resulting landslide susceptibility maps can be utilized for various land use management purposes.
6The study area is in the Gilgit-Baltistan region in the Central Karakoram region to the Northern part of Pakistan (fig. 1). This region was previously known as the Northern area. Furthermore, geographically, it is situated along the border, neighboring China to the Northeast and Afghanistan to the Northwest. The study area covers an elevation range of 1746 m to 7315 m above sea level and encompasses an area of 14,305.07 km2. This region is distinct due to its exceptionally high relief and harbors various hazards and topographic processes in the harsh environment of northern Pakistan. Topographically, the area covers high mountains with glaciers and tall peaks adorned with snow caps throughout the year. Geologically, particularly, a tectonic perspective, the area is greatly impacted by faults oriented in the NW-SE direction. Additionally, the valleys in this region have been shaped by glaciation and are abundant in sedimentary deposits from the late Quaternary and Holocene periods (Rehman et al., 2020). These deposits include unsorted glacial materials, debris flows, rock avalanches, and river gravels, particularly along the Hunza River (Seong et al., 2008). It is important to note that these sedimentary deposits can pose significant secondary hazards.
7Fig. 1 – Study area map showing provincial boundary, landslide location, and international boundaries.
Fig. 1 – Carte de la zone d'étude avec les limites provinciales, l'emplacement des glissements de terrain et les frontières internationales.
1. River; 2. Internationale boundary; 3. Provincial boundary; 4. Landslide; 5. Hunza watershed.
1. Cours d’eau ; 2. Frontière ; 3. Limite provincial ; 4. Glissement de terrain ; 5. Bassin versant de Hunza.
8The adopted methodology comprises six main steps: (i) constructing a detailed landslide inventory map of the pertinent study area, preparing using earlier records, satellite imageries, and detailed and careful field investigation; (ii) ten landslide conditional factors, Distance to Road, Aspect, Slope, Land cover, Stream Power index (SPI), Geology, Distance to Fault, Curvature, Distance to River, Land cover, and Cumulative rainfall data are selected after a detailed field survey of the pertinent study area; (iii) weight estimations for three statistical models including the proposed model and their results are compared for validation; (iv) the landslide susceptibility maps are prepared using three state-of-the-art deep machine-learning techniques and three statistical models; (v) the susceptibility maps produced from machine learning techniques will be validated using the F1 Score, Confusion matrix, Precision, and IOU curve; (vi) comparative analysis of Landslide susceptibility maps from all six models under the Seed Cell area Index SCAI, AUC, and Receiver Operating. See the schematic diagram shown in Figure 2.
Fig. 2 – Schematic flow diagram for the adopted methodology.
Fig. 2 – Schéma de flux pour la méthodologie adoptée.
The flow diagram shows the adopted methodology, landslide causative factors, models, and systematic steps.
Le diagramme montre la méthodologie adoptée, les facteurs causals des glissements de terrain, les modèles et les étapes systématiques.
9LIM map consists of historical pre-existing landslides and for reporting and summarizing the identified landslide occurrences worldwide, this map was later transferred to the database management system (Gorsevski et al., 2016). The LIM map provides detailed information about the location, intensity, and frequency of landslides in a particular area, which can help identify vulnerable infrastructure and mitigate the impact of future landslides (Abbas et al., 2023). Therefore, developing accurate and updated LIM should be a priority for risk management planning in landslide-prone areas before constructing a landslide susceptibility map. In this study, for the accompaniment of the precise data; the locations of landslide sites were identified with the assistance of Geologists and subject specialists after a detailed field survey over several weeks. The locations were marked using the Garmin GPS with an accuracy of 3 m in study areas. About 148 landslides and rock collapse sites were identified and marked in the pertinent study area. Once sufficient data has been collected, all information is digitized using Arc (GIS 10.2) software, which enables us to visualize the data for further analysis (fig. 3).
Fig. 3 – Landslide inventory map of the study area.
Fig. 3 – Carte d'inventaire des glissements de terrain de la zone d'étude.
A: acquired landslide points and locations; B: elevation map; C: geological map.
1. Landslide; 2. Faults ; 3. Stream.
A : emplacements des glissements de terrain acquis ; B : carte altimétrique ; C : carte géologique.
1. Glissement de terrain; 2. Faille ; 3. Cours d’eau.
10Identifying the landslide causative factors is significant to understanding the mechanism of prior landslides and their effects, providing temporal and spatial estimations on future landslides. These causative factors are severally temporal and spatially reliant on the geomorphology of the pertinent study area. Thus, it is essential at the initial phase to prioritize those factors to fit them in the study area. The GIS technology permits us to visualize these causal factors (Chen et Chen, 2021). Therefore, 10 landslide conditional factors are selected based on the critical review of previous literature and a detailed field survey. These selected conditional factors are Slope angle, distance to road, slope curvature, slope aspect, distance to fault, geology, distance to the river, stream power index (SPI), precipitation, and land cover. All the geomorphological, geological, and environmental conditioning factors are converted to thematic maps prepared in the ArcGIS 10.2 environment to build a geo-database for susceptibility analysis. The spatially disseminated thematic layers were projected using UTM Zone 43N to a 12 m x 12 m pixel resolution (fig. 4).
Fig. 4 – Thematic layers applied for landslide susceptibility modeling.
Fig. 4 – Couches thématiques appliquées à la modélisation de la susceptibilité aux glissements de terrain.
Thematic layers of 10 landslide conditional factors: A: slope aspect; B: curvature; C: distance to river (m); D: distance to road (m); E: distance to fault; F: geology formation; G: precipitations (mm); H: pentes (degrees); I: LC classes; J: SPI.
Couches thématiques de 10 facteurs conditionnels des glissements de terrain : A : expositions ; B : profil ; C : distance au cours d’eau (m) ; D : distance à la route (m) ; E : distance à la faille ; F : formation geologique ; G : précipitations (mm) ; H : pentes (degrés) ; I : classes LC ; J : SPI.
11Fuzzy set theory was developed by (Zadeh, 1965), and is commonly used to deal with the ambiguities of their membership in function. IF-D is a mathematical model used to measure the degree of similarity between two intuitionistic fuzzy sets. In fuzzy set theory, each element in the member has a value between 0 and 1, with 1 representing presence and 0 representing absence, and the degree of membership value represents the degree of relationship. Probability theory uses statistical divergence measures to measure the difference between two probability distributions (Kullback et Leibler, 1951). Moreover, among prominent theoretical divergence fuzzy models, intuitionistic Fuzzy Jensen-Renyi divergence (IFJRD) is one of the updated models used for probability analysis. While IFJRD provides more precise results than classical fuzzy models in probability analysis, this updated IFJRD model has too lengthy calculations and holds some limitations. This updated IFJRD may not adequately accommodate certain pattern sets (Verma et Sharma, 2013). Therefore, the research proposes a new (IF-D) model that performs better performance and caters to specific pattern sets in probability analysis landslide probability analysis (LSM). Therefore this IF-D divergence model is based on simple exponential divergence calculations which introduced the following algorithm to solve the probability of spatial distribution of landslides by using straightforward calculations.
12Step 01:
13Step 03:
14Ideal Solution
15Where stands for membership and stands for non-membership
16Step 04:
17Calculate
18Step 5:
19Where Dn stands for divergence, S(A, B) stands for similarity between A and B set ( landslides and non-landslide)
20The CNN architecture was improved to enable accurate segmentation of targets with few training samples. U-net has proven efficient in remote sensing segmentation tasks with three channels of images in the input. However, for LSM, 10 channels/bands of landslide conditional factors are used as input, requiring additional layers to the base model. Resnet34 is used in the encoder part to extract multi-scale features from the input data and remote sensory data, and max pooling performs downsampling in the encoder part with a stride of 2. In the decoder part, up-sampling is performed by increasing spatial size using bilinear interpolation. Pre-trained Image Net weights are used for the encoder, and additional convolutional layers, such as Batch Normalization and dropout layers at a rate of 0.2, are added to avoid over-fitting. The final convolutional layer has five neurons, and all convolution layers use the ReLU activation function, with the final output layer using SoftMax to provide the probability of a pixel belonging to a particular LSM class.
21InceptionV3+ is the latest state-of-the-art DMLTs Convolutional Neural Network (DM-NN) for image classification and belongs to the Inception family. It is the latest version of the basic model InceptionV1 and was developed by GoogleLeNet as the third series release in the Deep learning evolutionary Architectures (Szegedy et al., 2016). The outline of factorization in V3 objectives is to reduce the number of connections and parameters without banding the network's efficiency. The model comprises various building blocks such as convolutions, pooling, concerts, dropouts, and fully related layers, and improvements such as label smoothing, factorized 7 x 7 convolutions, and an auxiliary classifier have been made.
22DeeplabV3+ is the up-to-date state-of-the-art machine learning model with an encoder-decoder-based semantic segmentation network. The DeeplabV3+ has a setup of deeplabV3 architecture as an encoder at the backup of ResNet. The architecture is divided into three parts: the encoder used for feature extraction, ASPP to convert them into a wide scale information, and the encoder part arrayed to recuperate the spatial information. The extracted features in the encoder level are binary up-sampled and then concatenated with the respective low-level features in the subsequent stage. This model is constructed on Atrous Convolution and Atrous Spatial Pyramid Pooling (ASPP). In the Atrous convolution, the active field of view of the convolution is governed by a rate parameter. This Atrous convolution can be generalized as follows.
23Where w represents the filter, each location of output y, x is the input feature map and r denotes the Atrous rate.
24For the investigation, all factors contributing to landslides were initially categorized into a binary pattern (presence or absence of landslides) based on contrast weights, FR values, and divergence values. Categorical data such as lithology, land cover, precipitation, curvature, SPI, and aspect were then converted into binary patterns using weights calculated by WofE, FR, and IF-D for each factor class, with weights computed separately. The resulting weights were then used to assess the importance of each landslide conditional factor class in influencing landslide occurrences. Continuous data such as slope, distance to faults, distance to rivers, and distance to roads were classified into different categories, cumulative weights were calculated using the WofE, FR, and IF-D models for each factor class. The results of these weightings were also used to evaluate the importance of each landslide conditional factor class in influencing landslide occurrences. In this context, cutoff values were identified where the influence of factor classes on landslide occurrence was no longer statistically significant. The proposed model proved effective in determining these cutoff values for continuous data.
25Cumulative weights of Slope. The weight contrast (C), frequency ratio (FR), and divergence value (D) for each class of all landslides conditional factors were calculated using three bi-variate statistical models. The weight of the slope gradient was calculated in descending order if the influence of slope decreases as the slope gradient decreases. For the slope gradient classes, the highest values of FR and WofE, the high value indicated the highest probability of landslide occurrence was observed in the class range of 30-40° (FR = 1.62, C = 0.48). In comparison, the weight values of IF-D, and the cutoff value indicated the slope gradient below 20° and above 50° indicate does not influence landslides occurrence. Therefore, the cutoff value identified for the slope gradient was found 50°. Additionally, the minimum value of divergence D = 0 and maximum valve of S = 1 for the slope gradient 30-40° indicates the maximum probability of landslide occurrence (fig 5).
Fig. 5 – Graphical representation of weights derived from IF-D indicating cut-off values for continuous data.
Fig. 5 – Représentation graphique des poids dérivés de l'IF-D indiquant les valeurs seuils pour les données continues.
Cut-off values : A: for slope; B: for distance to fault; C: for distance to River; D: for distance to Road.
Valeurs seuil : A : pour la pente ; B : pour la distance à la faille ; C : pour la distance à la rivière ; D : pour la distance à la route.
26Cumulative weights of Distance to Fault. For the continuous data like distance to fault, the weight values of FR and WofE showed similar results. The values of FR are almost higher from 0 to 500 m but highest at 3000 m. Similarly, for WofE the contrast value is maximum for the class 3000-4000 m. Comparing the weight estimation from IF-D reveals that the maximum (D = 0, S = 1) for the class 3000-4000 m indicates maximum influence on landslide occurrence. However, the D value remains almost similar from 0 to 5000 m and the cutoff value was indicated below 5000 m (fig. 5, tab. 1).
27Cumulative weights of Distance to River. In the case of IF-D weight results for the influence of river on landslides occurrence, the IF-D values show shows (D = 0, S = 1) for the class 0-200 m and has a significant ascending trend but near to minimum value D = 0.02 for the class 600-800 m. This indicates that the maximum probability of landslide susceptibility regarding the effect of the river on landslide locations is between 0-800 m distances. Regarding the landside susceptibility association of the factor classes using WofE and FR, the results reveal that the strongest positive association (W+) appears the same for the class ranges from 0-800 m class. The overall results indicate that the maximum influence of distance to the river on landslide occurrence is between the range of 0-800 m (fig. 5, tab. 1).
28Cumulative Weights of Distance to Road. For the continuous data like road distance, the weight values of IF-D, and the value of divergence (D = 0, S = 1), the results reveal that the highest susceptibility of landslides occurrence is between the range of 0-500 m from the distance to the road. Additionally, the ascending divergence values from (0 to 0.44) and descending trend in similarity values (1.00 to 0.56) for the class 0-2000 m indicated that there was a significant influence of distance of road for the ranges 0-2000 m. However, based on the intersection point and support of the model the cutoff values were assigned up to 2000 m. In comparison with IF-D, the FR = 4.00 and C = 2.86 indicated the maximum influence of distance to the river on landslides is 500-1000 m. Therefore, the result estimated from IF-D is considered more effective as compared to FR and WofE (fig. 5, tab. 1).
29Categorical Weighs of Aspect. The multi-class generalization of continuous landslide conditional factors concerning the class borders was identified before the weights were calculated categorically. The second step was to summarize the weights for categorical data. The weights of WofE and FR indicate that (Flat, North, and N-E) have a rational relationship with landslide occurrence, particularly for the aspect class flat (WofE = 0.63, FR = 1.87). Moreover, the weights estimated from IF-D represented by (D = 0) indicate the maximum probability of landslide occurrence for the class flat aspect (tab. 1).
30Categorical Weighs of Curvature. In the WofE weight calculations, the shape of slope curvature has a positive association with landslide occurrence, with weights of 0.03 and 0.02 respectively. In addition, the contrast C = 0.05 indicates the maximum probability of landslide occurrence for the slope shape flat representing the maximum value. In comparing the weights evaluated from IF-D regarding the slope shape, the value of (D = 0) indicates the maximum probability of landslide occurrence is at flat shape class for slope curvature (tab. 1).
31Categorical Weighs of Lithology. The most susceptible class of lithology indicated by WofE and FR, the result showed the southern Karakoram metamorphic (C = 1.10, FR = 2.50), Quaternary deposits (C = 2.75, FR = 3) and Cretaceous sandstone (C = 1.32, FR = 3.7). These lithological units are very close to the active Klik thrust in the northern Pamir zone and the Main Karakoram thrust fault (MKT) in the South, therefore these lithological units are prone to landslides in the region. Similarly, weight estimated from IF-D revealed similar results, the southern Karakorum metamorphic, cretaceous sandstone, Permian massive, and quaternary deposits have the maximum values, indicating a high probability of landslide occurrence (tab. 1).
32Categorical Weighs of Stream Power Index (SPI). Regarding SPI, the analysis showed that the class (1.2 – 10.43) has the highest positive association (W+ = 0.45) for landslide occurrences based on WofE weight analysis. In addition, the class (-5.7 – -1.23) has the maximum value of S = 1 and minimum value of D = 0 based on IF-D, indicating the highest probability of landslide and the highest influence of SPI on landslide occurrence in this class.
33Categorical Weighs of Land Cover (LC). The predictive analysis indicates that orchards, settlements, summer pastures, and barren land have a positive relationship with landslides, with orchards having the highest weight-positive value (W+ = 1.08). The IF-D indicates that the highest probability of landslide occurrence is on barren land, which is prevalent in the study area due to the lack of natural forests and most of the slopes are barren, with settlements and irrigated lands below water channels. Landslides occurrence in the study area are mainly caused by channel excavation. Therefore, the derived values indicate most of the probability of landslides occurs on barren land (tab. 1).
34Categorical Weighs of Precipitation. In WofE weight calculations, class (18 – 24) showed a positive association with landslide occurrence, with W+ = 1.1 and contrast C = 3.4 indicating the maximum probability of landslide susceptibility. The FR weight estimation also showed a high susceptibility of landslides in this class, with the highest value of FR = 2.9. The IF-D model indicated the maximum probability of landslides occurrence for the class 18-24mm of precipitation, with divergence D = 0 (tab. 1).
Tab. 1 – Derived weight calculations using IF-D, FR and WofE..
Tab. 1 – Calculs de poids dérivés à l'aide de IF-D, FR et WofE .
35The machine learning models are subtle to data within their anticipated range. For creating DMLT models, numerical and categorical values of landslide conditional factors were prepared to produce landslide susceptibility maps using KERAS and GIS software. KERAS, built on top of TensorFlow, is a powerful Python library for building and training deep learning models, allowing easy deployment. In this study, the dependent factors were expressed as a binary variable (landslides and non-landslides) and normalized to a range of 0 to 1. To avoid overfitting, the dataset was divided into 70:30, with 70 % for training and 30 % for validation. Both negative and positive data were equally used to produce an LSM map. Validation is critical for evaluating DMLT models' performance. Validation tests such as the F1 score, IOU curve, and ACC score were calculated to confirm the models' significance and performance.
36The validation dataset was used to evaluate the performance of the DMLT models. Precision, recall, F1-score, and accuracy metrics were calculated, with the deeplabV3+ model achieving scores of 0.85, 0.89, 0.89, and 0.89 respectively. For IncetionV3+ the values of Precision, recall, F1-score, and accuracy are (0.93, 0.92, 0.92, and 0.92; fig 5). Similarly for U-Net, the values of Precision, recall, F1-score, and accuracy (0.85, 0.89, 0.89, and 0.89; tab. 2). Additionally, the models have been validated using the IOU curve and loss curve, the results revealed from the IOU for DeeplabV3+, InceptionV3+, and U-Net are (0.7,0.8, 0.6) and loss of (0.2, 0.3, 0.4; tab. 2). The results revealed that the models performed satisfactorily, and the models signify excellent prediction accuracy, hence results indicated that all the executed models performed well for the given data of the pertinent study area.
Tab. 2 – Accuracy statistics for deep machine learning models.
Tab. 2 – Statistiques de précision pour les modèles d'apprentissage profond.
37The final landslide susceptibility maps were created using three statistical models. Weight calculations were performed for each model, and the landslide conditional factors of each thematic layer were overlaid. The weights given to each factor were used to predict landslides when all factors were considered together. The final maps were prepared based on these weights, with geospatial datasets transformed into a raster format (12 x 12m grid). The factors were combined using the raster calculator tool in ArcGIS 10.2 to assess their impact on the final maps. For DMLTs models, susceptibility maps created using DeepLabV3+, InceptionV3+, and U-Net in the KERAS environment were transferred to GIS and divided into susceptibility classes (Very high, high, medium, and low). The natural break method, which maximized the area percentage of landslides in high and very high susceptibility classes, was used to examine landslide distribution in each zone of the created map (fig. 6).
Fig. 6 – Landslide susceptibility maps using statistical and DMLTs techniques.
Fig. 6 – Cartes de susceptibilité aux glissements de terrain utilisant des techniques statistiques et DMLTs.
Landslide susceptibility map using: A: DeeplabV3+; B: FR; C: IF-D; D: InceptionV3+; E: U-Net. F: WoFE.
Carte de susceptibilité aux glissements de terrain utilisant : A : DeeplabV3+ ; B : FR ; C : IF-D ;D : InceptionV3+ ; E : U-Net ; F : WoFE.
38Accuracy assessment of the map is a critical step in the overall process. Previous studies have applied various statistical techniques to validate the ability of a landslide susceptibility map, including the prediction rate curve (PRC), landslide density analysis (LDA), and seed cell area index (SCAI) (Kulsoom et al., 2023). Among these procedures, AUC is chosen due to its threshold independence and ability to measure accuracy and error rate. AUC has been widely used by multiple researchers for validating generated maps (Khan et al., 2022). Hence, this study employed AUC, LDA, and SCAI to assess the predictive accuracy of our map. For the training data set (70%) the statistical model exhibited the success rate curve (SRC) for WofE, FR, and IF-D are (82 %, 80 %, and 75 %) respectively (fig. 7D). Likewise for the prediction rate cure the statistical models of the prediction rate cure (PRC) showed WofE, FR, and IF-D are (85 %, 82 %, and 85 %) respectively (fig. 7C).
39Similarly, for DMLTs on the training data set (70 %) the value of (SRC) showed for DeeplabV3+, InceptionV3+, and U-Net are (79 %, 75 %, and 78 %) respectively. Similarly, from the validation data set (30 %) the prediction rate curve (PCR) the value for DeeplabV3+, InceptionV3+, and U-Net are (82 %, 80 %, and 79 %; fig. 7A, B).
Fig. 7 – Prediction accuracies and comparative assessment of prediction rate for all models.
Fig. 7 – Précisions de prédiction et évaluation comparative du taux de prédiction pour tous les modèles.
A : courbe du taux de prédiction (PRC) pour les modèles DMLTs ;B : courbe du taux de succès (SRC) ;C : courbe du taux de prédiction (PRC) pour les modèles statistiques ;D : courbe du taux de succès (SRC) pour les modèles statistiques.
40In this study, the landslide susceptibility map (LSM) is produced in the ArcGIS 10.2 environment while the DMLTs executed in the KARES python programming are then exported to ArcGIS 10.2 creating LSMs. The final landslide susceptibility index map was classified with 4 levels of classification: 1-Very low, 2-Medium, 3-High and 4-very high.
41In the landslide susceptibility map generated using statistical models, the landslide density (LDA) was observed to be 84.8 % for WofE in the high and very high susceptibility class, 88 % for FR, and 83.6 % for IF-D (fig. 8A). In the case of DMLT models, the LDA was determined to be 79.5 % for U-net in the high and very high susceptibility class, 76 % for DeeplabV3+, and 3.4 % for inceptionV3+. Among all models, InceptionV3+ showed less precision and prediction of the distribution of LDA in each susceptibility zone (fig. 8B).
Fig. 8 – Distribution of landslides in each susceptibility zone for all models.
Fig. 8 – Répartition des glissements de terrain dans chaque zone de susceptibilité pour tous les modèles.
A: the distribution of landslides in each susceptibility zone for bi-variate statistical models; B: the distribution of landslides in each susceptibility zone for DMLT models.
A : répartition des glissements de terrain dans chaque zone de susceptibilité pour les modèles statistiques bivariés ; B : répartition des glissements de terrain dans chaque zone de susceptibilité pour les modèles DMLT.
424.8. Validation and Comparison of LSM Based on SCAI
43To analyze the classification capacity of six LSM models more accurately, Seed Cell area Index (SCAI) tests were applied which explain the differences between the divided zones. The larger the difference between the divided zones, the larger the classification capacity of LSM. This study used the Seed cell area index (SCAI) to estimate the differences among six LSMs (Suzen et Doyuran, 2004). The landslide grid cell is called a “Seed cell” and the SCAI can be calculated using the following equation.
44% 𝑎𝑟𝑒𝑎 indicates the percentage of grid cells in each susceptibility class to total grid cells in the entire area while % 𝑠𝑒𝑒𝑑 indicates the percentage of landslide grid cells in each susceptibility class to grid cells of all landslides. The SCAI values for wofE, FR, and IF-D were 4.3, 4.9, and 2.4 respectively. Likewise, the SCAI values for DMLTs were 4.8, 8.8, and 4.7 for DeeplabV3+, InceptionV3+, and U-Net respectively, see fig 9A. However, the difference in SCAI was indicated by the D-value. Including all six LSM models, the D-values were 2.9, 3.8, and 2.4 for WofE, FR, and IF-D respectively. For the DMLTs models, the D-values were 3.6, 3.0, and 3.3 for DeeplabV3+, InceptionV3+, and U-Net (fig. 9B).
45Landslide susceptibility assessment has been the emphasis of various techniques in many research papers over the past two decades. Between 2005 and 2019, 70 statistical models were applied for LSM in different landslide-prone areas worldwide. The Frequency Ratio (FR) method was applied in 92 cases, and the Weight-of-Evidence (WofE) method in 60 cases (Pourghasemi et Rossi, 2017). These two statistical models have seemed as the most widely used and prevalent methods in the past few decades (Ozdemir et Altural, 2013). The combination of frequency ratios (FR) and weights of evidence (WofE) provides multiple benefits for conducting landslide susceptibility assessment (Qazi et al., 2023). The use (of WofE) and (FR) methods in landslide susceptibility modeling have limitations, particularly for examining continuous variables in recognizing cut-off values. This can introduce unfairness and compromise the accuracy of the resulting susceptibility maps (Khan et al., 2022). To address these limits, a substitute method was used to accurately estimate both categorical and continuous data a model (IF-D) has been proposed, and its results were compared with conventional models (FR et WofE) the results showed the proposed model has straightforward for estimation of cutoff values with precise estimations (fig. 5).
46Furthermore, over the past decade, of deep-learning models, there has been inadequate literature available on comparative studies for landslide susceptibility analysis (Zhang et al., 2021). However, Deep learning techniques have extraordinary feature learning and classification capabilities that have led to many imposing accomplishments (Niu et Chen, 2019). DeepLabV2 has emerged as a leading framework in this domain due to its extraordinary performance (Chen et al., 2018). Additionally, the U-Net was successfully applied for the first time in glacier risk assessment in the Karakoram terrain. The estimation of the test data revealed that the U-Net achieved an impressive F1 Score of 0.936, further strengthening its efficacy in this domain (Qayyum et al., 2020). In addition, in this research, the author has presented more cutting-edge and upgraded DMLT models such as; DeepLabV3+, InceptionV3+, and U-Net for landslide susceptibility modeling in Karakoram terrain.
47The study aimed to evaluate landslide conditional factors, their spatial distribution, and their influence on landslide occurrence by comparing the proposed IF-D model with the Weight of Evidence (WofE) and Frequency Ratio (FR) methods. The Intuitionistic Fuzzy Divergence (IF-D) method was compared to WofE’s maximum contrast value (C) and FR's maximum value for validation purposes. Two methods, WofE using studentized contrast (Cs) and IF-D were comperatively examined to determine the cut-off value for continuous data.
48However, assessing slopes is of particular utility in understanding past slope failure. A statistical calculation has been carried out to evaluate the influence of slope on landslide occurrence. For the continuous data like slope, the weight values gained from both approaches indicated that slope gradients below 20° and above 60° had negligible impact on landslide occurrence, see Appendix A. However, the overall statistical weight estimations revealed that slope gradients ranging from 30° to 40° strongly influence landslides in the study area (tab. 1).
49Similarly, geological faults are significant contributing factors for landslide occurrence. In the study area, regional faults to the south and a local fault system to the north play a leading role in triggering landslides (Ali et al., 2017). The main Karakoram thrust (MKT) and the Karakoram Fault are important structural features in the study area (Ali et al., 2019). Furthermore, from the statistical results, the impact of faults on landslide occurrences is found within a range of 1000 to 3000 m. In the weight estimation using the IF-D model, it was observed that the class 3000-4000 m has the highest D values, indicating a significant impact on landslide occurrence. However, the D value remains relatively stable within the range of 0 to 5000 m, with a cut-off value below 5000 m (tab. 1).
50Additionally, during the construction of the KKH in the study area, slope cuttings and vibrations from heavy machinery enlarged landslide vulnerability. The (WofE, C = 2.86) and (FR = 4.00) exhibited that the 0-500 m distance to the road had the greatest influence on landslide occurrence. Similarly, IF-D weight estimation yielded similar results, with a minimum divergence (D) of 0 indicating the highest influence within the 0-500 m range (tab. 1).
51Additionally, from the analysis of categorical data like slope curvature, the WofE et IF-D, the weighting analysis indicated that both "Concave" and flat classes have the maximum influence on landslide occurrence. This result was then compared to the inventory map, which exposed that the overall findings from WofE and IF-D estimate the weight estimation more precisely than the FR Model (tab. 1).
52A statistical assessment was performed to evaluate the spatial influence of lithologies on landslide occurrences. In comparison to FR and WofE, the weighting values of IF-D yielded similar and accurate results for the southern Karakoram metamorphic (D = 0.02) Permian massive (D = 0.07), Cretaceous sandstone (D = 0.28), and Quaternary deposits (D = 0.00, S = 1), but the weight derived from IF-D analysis established that Misgar slates (D = 0.24) had a substantial impact on landslide occurrences. Also, the weight derived from IF-D analysis highlighted the significance of Misgar slates on landslide occurrences. The results were confirmed through a comparison of the derived weights with the inventory map. It was found that the derived weights from the (WofE) and (FR) methods yielded vague values regarding the Misgar lithological uni (tab. 1).
53The comprehensive statistical valuation, encompassing the employment of FR, WofE, and the newly proposed IF-D model, determined that weight estimation for continuous data was consistently accurate across all three methodologies. Furthermore, upon juxtaposing the weight outcomes with field observations and inventory maps it became evident that the weight estimates derived for categorical data through the utilization of the proposed IF-D model exhibited a higher degree of precision when contrasted with traditional models WofE and FR.
54Landslide susceptibility assessment involves statistical analysis to correlate landslide causative factors with landslide occurrences, applying a bi-variate statistical approach and a refined model called Intuitionistic Fuzzy Divergence (IF-D). The IF-D model improved over traditional fuzzy models and produced more accurate results. Comparative weight estimation for continuous and categorical data revealed that traditional models (FR et WofE) produced inaccurate estimations for categorical variables, while the IF-D model provided more accurate results. To create final landslide susceptibility maps, weights were given to each thematic layer for each statistical model, integrating ten landslide conditional factors. The Landslide Susceptibility maps were classified based on the natural breaks method, and the model goodness of fit was evaluated using, accuracy based on the seed cell area index (SCAI), and prediction accuracy (AUC). Consequently, WofE and IF-D appeared as the most appropriate and reliable models for landslide susceptibility mapping in the study area, with statistical models overall outperforming deep learning models in effectiveness and reliability. These maps may be useful for future land use planning, engineering purposes, and for the community living in these valleys.
*Corresponding author: Asghar Khan, Tel: (+92) 5811960010-1 (asghar.khan@kiu.edu.pk)