Extreme-rainfall landslide clusters are numerous, morphologically continuous and process-mixed, while event inventories mainly record boundaries and locations, limiting type and process interpretation. We develop a self-supervised framework for landslide-type recognition and probability mapping. Terrain point clouds encode landslide boundaries, internal topography and hydrogeomorphic context, and a terrain-context model transfers object-level type structure into spatially continuous probabilities. For the Wuping event in Fujian, China, 20,440 samples were interpreted as planar, flow-like, convergent and micro landslides. Except for micro landslides, the three main types were separable from terrain context, with differences mainly controlled by channel proximity, slope-flow paths and local relief. Terrain-response-unit mapping and 82 field photographs support post-event mapping, risk interpretation and field identification.

rainfall-induced landslides,self-supervised learning,terrain point cloud,probabilistic mapping,field identification