Improved Mask R-CNN inventory mapping and multivariable geo-environmental assessment of rainfall-induced shallow landslides in mountainous Beijing, China

Extreme rainfall can trigger numerous small, boundary-fragmented and spatially clustered shallow landslides. This study combined improved instance segmentation with reproducible spatial analysis to map and assess shallow landslides associated with the 2023 ‘23.7’ rainstorm in Zhaitang Town, mountainous Beijing. An improved Mask R-CNN integrated a convolutional block attention module, a path aggregation feature pyramid network and Soft-NMS. Under the random tile-level test design, the model achieved 89.51% precision, 88.31% recall, an F1 score of 88.91% and a mean intersection over union of 80.03%; these are in-scene estimates rather than evidence of geographic transferability. The resulting inventory contains 8298 landslides covering 3.05 km². Analysis of the harmonized geospatial layers showed that kernel density > 69.5 landslides/km² occupied 6.42% of the valid study area and contained 48.90% of mapped landslides. Frequency-ratio analysis identified enrichment in Wumishan Formation Members 1–3, 400–800 m elevation, 30–60° slopes, east-to-southwest aspects and concave terrain. A presence–background logistic model using 8212 landslide locations and 48,553 valid fixed-seed background samples showed no independent association for cumulative rainfall (odds ratio [OR] 1.00 per standard deviation, p = 0.999) or hourly intensity (OR 0.91, p = 0.407) after adjustment. Elevation, slope, plan curvature, aspect and lithology remained associated. Rainfall×slope and rainfall×susceptible-lithology interactions were non-significant. Residual Moran’s I was significant at 1 km but non-significant at 2–4 km aggregation, whereas five-fold spatial-block validation varied among regions (ROC AUC 0.699–0.870). The results separate event attribution from within-event spatial association and demonstrate a transparent workflow from rapid inventory mapping to spatially diagnosed environmental assessment.