Abstract
Landslides, as a natural hazard, pose a potential risk to human life and infrastructure. However, predicting and mitigating landslide susceptibility is challenging due to the complexity of contributing factors, such as slope, groundwater level, earthquakes, and vegetation. This study aims to develop a GraphSAGE–Graph Attention Network (GraphSAGE-GAT) and Best–Worst Method (BWM) for predicting landslides and generating landslide susceptibility maps (LSMs) in Savadkouh County (northern Iran). For this purpose, twelve conditioning factors relevant to LSMs were considered: slope, faults, precipitation, vegetation, soil, watercourses, roads, cities, villages, elevation, and temperature. Receiver operating characteristic (ROC) curves were used to evaluate model performance based on the number of events in the landslide susceptibility classes, as well as the Precision–Threshold curve, Recall, Precision, and Specificity across all possible thresholds. The most important factor in the BWM was slope, with 22.63%. By combining weights and layers in GIS, LSMs were generated using both BWM and GraphSAGE-GAT. The results show that 47% of the landslide points were classified as high susceptibility (high and very high) by BWM, whereas 67.33% were classified as high susceptibility by GraphSAGE-GAT. Furthermore, BWM achieved a maximum Precision of approximately 73%, while the Precision of GraphSAGE-GAT increased with the threshold and approached 100% at very high threshold values. The area under the ROC curve was 0.544 for BWM and 0.836 for GraphSAGE-GAT, demonstrating the substantially better discriminatory performance of GraphSAGE-GAT. Therefore, GraphSAGE-GAT integrated with GIS provides an effective approach for assessing landslide susceptibility, and the results are promising because the model not only provides good discrimination between susceptible and non-susceptible areas but also generates LSMs that capture spatial relationships among landslide-conditioning factors at the county scale.