High-resolution landslide detection dataset in the sparsely vegetated mountainous regions

Considering the heterogeneous land-cover conditions within the study area, which is a typical arid mountainous region characterized by sparse vegetation and complex surface exposure, representative non-landslide samples were carefully constructed to improve the robustness of the dataset for landslide instance segmentation. Specifically, non-landslide categories were systematically collected from areas surrounding documented landslide events, including urban settlements, transportation corridors, river channels, barren land, cropland, forests, and snow-covered surfaces. This design ensures a comprehensive representation of background diversity under arid and semi-arid environmental conditions, thereby reducing class imbalance and enhancing the model’s ability to distinguish landslide objects from visually similar non-landslide terrains. In addition, to incorporate physically meaningful slope stability information into the proposed landslide instance segmentation framework, we developed an automated steady-state factor-of-safety (FS) computation pipeline for regional-scale slope stability assessment. The developed toolchain is publicly available at https://github.com/syty10/landslide_fs_tools, enabling efficient generation of FS-based susceptibility priors from user-provided shapefile (SHP) data with minimal preprocessing requirements.