1. Introduction

Landslides represent one of the most destructive natural hazards worldwide, causing substantial economic losses and human casualties annually. These mass movements result in the destruction of forests, fertile agricultural lands, residential areas, communication networks, and tourism infrastructure. Furthermore, landslides contribute to severe geomorphic modifications and may trigger secondary hazards such as flooding in hilly regions. Iran, due to its unique physiographic, environmental, and climatic conditions—coupled with anthropogenic pressures and increasing demands on natural resources—is highly vulnerable to landslide occurrences, particularly in its mountainous areas. The country experiences various natural threats, including severe soil erosion, devastating floods, and destructive landslides, which have collectively resulted in billions of Rials in financial damages. Consequently, landslides have become a national disaster in Iran, necessitating systematic approaches to assessment and management. The Zagros mountain range, where the study area is located, is particularly prone to mass movements due to its complex geological structures, high seismicity, and seasonal heavy rainfall. Previous studies have reported numerous landslide events along the main roads and rural settlements in this region, causing annual economic losses estimated at several million dollars. Despite these challenges, systematic landslide risk assessments at the watershed scale remain limited in Lorestan Province. Most existing studies have focused only on susceptibility mapping without considering elements at risk and their vulnerability. This gap highlights the urgent need for comprehensive risk assessment frameworks that integrate all three components of risk (hazard, elements at risk, and vulnerability) to support effective land-use planning and disaster risk reduction strategies.There is no universally accepted method for landslide susceptibility mapping due to the wide range of influencing factors and local environmental conditions. However, the application of logical and quantitative approaches, including statistical and machine learning models, has significantly improved the accuracy and reliability of landslide zoning maps. In recent years, data mining and machine learning techniques have gained widespread attention due to their high predictive accuracy and robust information processing capabilities. Among these, Artificial Neural Networks (ANN), Generalized Linear Models (GLM), and Support Vector Machines (SVM) have demonstrated excellent performance in landslide susceptibility zoning. While landslide susceptibility assessment focuses on the spatial probability of landslide occurrence, a comprehensive landslide risk assessment requires the integration of three interrelated components: landslide hazard (probability of occurrence), elements at risk (e.g., roads, water sources, buildings, agricultural lands, power transmission lines), and vulnerability of those elements. The present study aims to address existing gaps by conducting a comprehensive landslide risk assessment in the Varosht Watershed, Lorestan Province, Iran, using three machine learning models.



2. Materials and Methods

The Varosht Watershed is located in northwest Lorestan Province, Iran, within geographical coordinates of 33°06′18″ to 33°58′19″ North latitude and 47°50′08″ to 48°41′09″ East longitude, covering a total area of 15,435 hectares. The minimum elevation is 1,509 meters above sea level, while the maximum elevation reaches 2,620 meters. Approximately 75% of the watershed area is characterized by slopes exceeding 15%, indicating the mountainous nature of the study area. Due to its steep slopes, susceptible geological formations, road construction activities, and widespread land use changes—particularly the conversion of forests and rangelands into rain-fed agricultural lands—the Varosht Watershed exhibits high susceptibility to landslide occurrence. The mean annual precipitation is 469 mm. The precipitation pattern follows a Mediterranean climate regime, with most rainfall occurring between November and March. This seasonal concentration of rainfall, combined with the region's steep topography and weathered geological formations, creates favorable conditions for landslide initiation, particularly during periods of intense or prolonged rainfall. Additionally, the study area has experienced several moderate to strong earthquakes in the past century, which have further weakened the rock mass and increased the overall slope instability.According to Iran's structural-tectonic divisions, the study area is situated within the Crushed (High) Zagros structural zone, with predominant lithology consisting of limestone, conglomerate, marl, radiolarite, and alluvial deposits. A landslide inventory map was prepared based on 163 landslide polygons systematically surveyed and recorded. The total landslide-affected surface within the watershed is approximately 131.32 hectares. The smallest documented landslide covers 128 m², while the largest occupies 13.32 hectares, demonstrating considerable heterogeneity in landslide magnitude across the study area. Seventeen conditioning factors were selected as independent variables influencing landslide occurrence: slope percentage, aspect, elevation, landform, length-slope (LS) factor, Topographic Position Index (TPI), Topographic Roughness Index (TRI), Vector Ruggedness Measure (VRM), Topographic Wetness Index (TWI), distance to faults, distance to roads, distance to villages, lithology, land use, Normalized Difference Vegetation Index (NDVI), mean annual rainfall, and distance to streams. Primary derivatives were extracted from a Digital Elevation Model (DEM), while advanced geomorphometric indices were computed using SAGA GIS software. These seventeen factors were selected based on a comprehensive review of the landslide susceptibility literature and the specific physiographic characteristics of the Zagros region. All factor layers were resampled to a uniform spatial resolution of 30 meters and standardized to ensure compatibility across different measurement units. The dataset was randomly split into training (70% of the data) and validation (30% of the data) subsets to ensure unbiased model evaluation. Three machine learning models were employed:

• Support Vector Machine (SVM): A supervised learning method that finds an optimal hyperplane for separating landslide and non-landslide data points in high-dimensional space.

• Generalized Linear Model (GLM): An extension of traditional linear models accommodating non-normal distributions, expressed as logit(P) = β₀ + Σβᵢxᵢ.

• Artificial Neural Network (ANN): A multilayer perceptron with one input layer, one hidden layer (ReLU activation), and one output layer (sigmoid activation), trained using backpropagation with Adam optimizer.

All models were implemented using ModEco software and validated using Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) metrics. Elements at risk included residential zones, transportation infrastructure (asphalt and unpaved roads), land cover types (forest, rangeland, agricultural land), surface water resources, and power transmission lines. Vulnerability scoring was conducted by integrating the inherent value of elements with hazard classes. The final risk map was produced using the equation R = H × E × V (Varnes, 1984), classified into five classes: very low, low, medium, high, and very high risk.



3. Results and Discussion

Validation results demonstrated that the SVM model achieved the best performance with an AUC of 0.913, followed by ANN (0.865) and GLM (0.787). The superior performance of SVM is attributed to its ability to find an optimal hyperplane for separating landslide and non-landslide data points, even in high-dimensional spaces. These findings are consistent with previous studies by Peng et al. (2014), Pham et al. (2016), and Chen et al. (2017a), which introduced SVM as one of the most efficient machine learning algorithms for identifying landslide-prone areas. According to the SVM model, approximately 35.66% of the area was classified as very low susceptibility, 22.62% as low, 17.09% as medium, 12.68% as high, and 11.94% as very high susceptibility. In contrast, the GLM model placed a higher percentage of the area (55.57%) in medium to very high susceptibility classes compared to the SVM model (41.71%), indicating the more conservative approach of SVM. The superior performance of SVM over ANN and GLM can be explained by several factors. First, SVM is inherently designed to handle high-dimensional feature spaces effectively, making it suitable for landslide susceptibility mapping where multiple interacting factors influence slope stability. Second, the use of kernel functions in SVM allows the model to capture complex nonlinear relationships between conditioning factors and landslide occurrences without suffering from overfitting. Third, SVM's reliance on support vectors (a subset of training data) makes it less sensitive to outliers and imbalanced datasets compared to ANN and GLM. In contrast, GLM assumes linear relationships between predictors and the log-odds of landslide occurrence, which may not hold true in complex mountainous environments. ANN, while capable of modeling nonlinear patterns, requires careful tuning of network architecture and is more prone to overfitting when training data are limited. The final risk map revealed that approximately 54.86% of the watershed falls within the very low-risk class, 26.08% in low-risk, 10.84% in medium-risk, 4.28% in high-risk, and 3.94% in very high-risk classes. Overall, about 19% of the area (equivalent to 2,968 hectares) is classified as high and very high-risk classes, necessitating prioritization in management and mitigation measures.

Comparison of elements at risk and vulnerability maps revealed that although the largest area of the watershed (approximately 59%) falls within the low vulnerability class, areas with high and very high vulnerability (totaling about 11%) largely overlap with high hazard classes. This overlap, reflected in the final risk map, emphasizes the importance of an integrated risk assessment approach combining hazard, elements at risk, and vulnerability.



4. Conclusion

This study draws several important conclusions:

1. The Support Vector Machine (SVM) model, with an AUC of 0.913, achieved the best performance among the three models examined and was selected as the optimal model for landslide susceptibility zonation in the study area.

2. The Artificial Neural Network (ANN) model (AUC=0.865) and Generalized Linear Model (GLM) (AUC=0.787) ranked second and third, respectively.

3. The final landslide risk map revealed that approximately 19% of the watershed area (2,968 hectares) falls within high and very high-risk classes, requiring prioritization in management measures.

4. The integrated risk assessment approach based on the equation R = H × E × V provides an effective method for identifying high-risk areas.

5. The seventeen conditioning factors used constitute a comprehensive set of variables influencing landslide occurrence.

6. The landslide inventory map with 163 polygonal landslides provided a robust foundation for calibrating and validating the machine learning models.

It is recommended that future studies employ deep learning models and ensemble approaches to further improve predictive accuracy. Furthermore, greater attention should be paid to casualty damage (human losses), which was not addressed in this study due to lack of appropriate data. The results of this study can serve as a scientific basis for land use planning, sustainable development, and landslide risk mitigation in the mountainous regions of Iran.