GeoBASP: Geography-Aware Structured Pruning for Landslide-Scar Segmentation in the Nepal Himalaya

Rapid landslide mapping is important for disaster response, yet modern segmentation networks can require substantial computation and storage. This study proposes Geography-, Boundary-, and Scar-Aware Structured Pruning (GeoBASP) for compact landslide-scar segmentation. GeoBASP measures the importance of removable channel groups from four complementary aspects: complete-mask accuracy, boundary delineation, small-scar detection, and the rejection of difficult non-landslide areas. These signals are combined across training blocks within the Koshi Basin using robust statistics to guide pruning under measured computational budgets. Experiments used a spatially separated dataset derived from the ICIMOD Koshi Basin 2024 inventory, Sentinel-2 imagery, and terrain slope. ResNet-50 U-Net, ConvNeXt-V2-Tiny U-Net, and SegFormer-B2 were pruned at 25%, 50%, and 70% multiply–accumulate (MAC) reduction. Two general-purpose baselines were evaluated across all three architectures and one segmentation-oriented method on the two U-Nets. GeoBASP achieved the highest overall mean validation Dice in the three-architecture comparison and exceeded the segmentation-oriented method at 25% and 50% reduction on both U-Nets. Rankings varied with architecture and compression level. Only GeoBASP satisfied the predefined non-inferiority condition, with qualifying runs restricted to the 25% target; no method satisfied it at 50% or 70% reduction. For the highest-performing dense architecture, adding slope to the imbalance-aware spectral input increased the mean test Dice from 0.471±0.007 to 0.520±0.004. At the moderate operating point, GeoBASP reduced MACs by 24.2% and parameters by 12.3% while retaining a test Dice of 0.505 compared with 0.516 for the dense model. For ConvNeXt-V2-Tiny U-Net, validation Dice remained 0.461±0.008 at approximately 50% MAC reduction and 0.419±0.004 at 70%, with corresponding parameter reductions of 33.3% and 53.7%. These findings support moderate compression within the evaluated basin, while performance depends on architecture and aggressive pruning incurs accuracy losses.