Scalable landslide detection in complex Himalayan topography via a novel object-based residual graph attention network

High-mountain environments are some of the most fragile and difficult environments to manage. The local communities are always under stress due to lack of sufficient land resources for their sustenance and frequent occurrences of landslides further complicate the situation. Landslide detection in high-mountains is increasingly critical as climate-induced deglaciation, permafrost thaw, and shifting precipitation patterns drive slope instability and cascading hazards. Timely mapping these events on satellite images remains challenging due to complex topography and spatial heterogeneity, while pixel-based deep learning often suffers from high costs and noise sensitivity. Here, we propose an open-source object-based Landslide Residual Graph Attention Network (LANDS-ResGAT) for landslide segmentation in rapidly transforming mountains. Superpixel-based segmentation generates spatial units, from which spectral and DEM-derived topographic features and terrain metrics are extracted. A spatial object graph encodes contextual relationships via centroid distances and edge attributes. LANDS-ResGAT’s 8-layer HybridResNet, integrating SAGEConv and GATv2Conv with residual connections, learns robust structural and contextual patterns. Evaluated on six Himalayan subregions with 3 m/pixel resolution PlanetScope satellite imagery, it achieves F1-scores above 0.92, outperforming Swin Transformer and U-Net++ while reducing inference time to under 3.5 min per scene. This scalable, interpretable approach can be adopted for any other high-mountain region globally to support real-time risk assessment and climate-sensitive geohazard monitoring.