Spectral-Confusion-Suppressed Landslide Recognition in High-Resolution Remote Sensing Images Using Geological-Feature Enhanced Cascade Mask R-CNN

Abstract Landslide disasters critically threaten human safety and infrastructure. This study proposes a geological-feature enhanced Cascade Mask R-CNN framework to resolve spectral confusion (e.g., bare rock misclassification) in 0.8-m resolution remote sensing images. First, a Geological Feature Pyramid Network (GFP-Net) integrates NDVI indices and texture features via gated weighting to amplify landslide signatures while suppressing irrelevant spectra. Second, a dynamic loss function adaptively adjusts weights by sample distribution, mitigating landslide/non-landslide class imbalance. Validated on the Bijie dataset (1,905 test images), our method achieves 91.7% precision, 89.6% recall, and 90.3% mAP — outperforming YOLOv8 by 6.6%/5.4%/7.4%. Notably, false positives reduce by 48.4% (48 vs. 93), with bare rock misclassifications dropping from 18.7% to 7.1%. The framework demonstrates superior boundary segmentation in complex terrains, providing an operational tool for automated landslide mapping and geohazard prevention.