Journal of Watershed Management Research

Introduction and Objective: Landslides are recognized as one of the most significant geological hazards, causing considerable damage to human life, infrastructure, and the environment every year, and imposing major economic burdens on governments. The Taleghan Dam basin, due to its mountainous terrain, high rainfall, favorable climate for construction, and unregulated road development, is considered one of the regions highly susceptible to landslide occurrences. Identifying and avoiding landslide-prone zones is among the most cost-effective measures that can be undertaken. In recent years, this has been made possible through the use of advanced machine learning models in conjunction with Geographic Information Systems (GIS) and remote sensing techniques, with researchers achieving promising results. The main objective of this study is to identify landslide susceptibility zones and prioritize the contributing factors using two machine learning models such as Artificial Neural Network (ANN) and Support Vector Machine (SVM) integrated within a GIS environment. According to the literature review, no comprehensive study has yet been conducted regarding landslide susceptibility mapping in the Taleghan Dam basin. Therefore, this research is essential for evaluating and managing the watershed to prevent and reduce damages caused by landslides.

Materials and Methods: To assess the spatial probability of landslides, a landslide inventory map of previously recorded events was first prepared using field surveys, local sources, and data from the Alborz Province Department of Natural Resources and Watershed Management, within the ArcGIS 10.8 environment. Eleven factors influencing landslide potential were selected, including elevation (meters), slope (percent), slope aspect, distance from streams (meters), lithology, precipitation, land use, Topographic Wetness Index (TWI), surface curvature, and distances from streams and roads. Multicollinearity among the variables was evaluated using the Variance Inflation Factor (VIF) test in SPSS software. Out of a total of 631 landslides, 70% (442 points) were randomly selected for training, and 30% (189 points) were used for validation. The SVM and ANN models were applied to identify landslide-prone areas. The Jackknife method was used to determine the most influential parameters, and the Receiver Operating Characteristic (ROC) curve was used to evaluate the predictive power of the models. The superior model was then selected based on ROC values, and the final landslide susceptibility map was produced.

Results: The results indicated no multicollinearity among the input factors, allowing all of them to be used in the modeling process. According to the Jackknife test, the layers representing distance from faults, slope aspect, distance from roads, lithology, and land use were identified as the most influential factors in landslide susceptibility, in descending order. The Area under the Curve (AUC) from the ROC analysis demonstrated model accuracies of 90% (excellent) and 88% (very good) in the training phase, and 87% (very good) and 84% (very good) in the validation phase for the ANN and SVM models, respectively. Based on the superior model and the final susceptibility map, approximately 64.5% of the watershed area was identified as having moderate to high landslide susceptibility. The “very low” susceptibility class covered the smallest area (12.3%), while the “moderate” class covered the largest area (27.3%). In terms of spatial prioritization, the sub-watersheds of Lambaran, Grap, and Zidasht 1 were found to have the highest landslide susceptibility, while the sub-watersheds of Nesasofla, Shahrak, and Khoranak exhibited the lowest.

Conclusion: The findings demonstrated that the application of machine learning techniques is not only cost and time-efficient but also highly capable of accurately predicting landslide susceptibility. Given the excellent accuracy of the produced susceptibility maps, the results of this study can serve as valuable tools for decision-makers, local authorities, and planners in order to reduce and manage landslide-related damages.