Spatially Explicit and Temporally Validated Graph Convolutional Network Framework for Landslide Susceptibility Mapping in Basaltic Terrain

Landslide susceptibility mapping (LSM) is important in rapidly developing, landslide‐prone basaltic terrains of the Western Ghats, India. Although temporal validation has gained attention in LSM, comprehensive comparisons of graph‐based deep learning and conventional machine learning models incorporating temporal and spatial validation, uncertainty quantification and reliability assessment remain limited. In this study, a spatially explicit and temporally validated framework is implemented to evaluate a graph convolutional network (GCN) against deep neural network (DNN), one‐dimensional convolutional neural network (1D CNN), random forest (RF), extreme gradient boosting (XGB) and light gradient boosting machine (LGBM) to better capture spatial relationships in the basaltic terrain of Raigad district. A dataset of 1840 samples (920 landslide and 920 non‐landslide) was prepared, with 1412 samples used for model development and 428 samples (214 landslide and 214 non‐landslide) from 2022 to 2025 reserved for testing. Model performance was assessed using multiple evaluation metrics supported by Taylor diagram analysis, stratified and spatial block cross validation and bootstrap‐based uncertainty analysis. GCN achieved the highest performance (AUCROC = 0.989; accuracy = 0.9239), followed by XGB (0.974), RF (0.970), LGBM (0.969), 1D CNN (0.886) and DNN (0.865). Permutation importance analysis identified distance to roads, slope, NDVI, relative relief and annual rainfall as dominant factors. The proposed framework enhances the reliability and spatial generalization of landslide susceptibility assessment, providing a robust basis for risk‐informed planning in data‐scarce mountainous regions.