Spectral-confusion-suppressed landslide recognition in high-resolution remote sensing images using geological-feature enhanced Cascade Mask R-CNN

Abstract

Landslide disasters pose critical threats to human safety and infrastructure. While high-resolution remote sensing imagery offers substantial potential for automated landslide recognition, two persistent challenges remain: spectral confusion between landslides and geologically analogous features such as bare rocks and barren surfaces, and severe class imbalance wherein landslide pixels constitute only a minor fraction of the total image area. This paper proposes a geological-feature enhanced Cascade Mask R-CNN framework to address these challenges. The framework introduces two key innovations: a Geological Feature Pyramid Network that integrates Normalized Difference Vegetation Index and texture descriptors through a gated fusion mechanism to amplify discriminative landslide signatures while suppressing irrelevant spectral information, and a dynamic loss function that adaptively adjusts weights based on batch-wise sample distribution to mitigate class imbalance. Extensive experiments conducted on the Bijie dataset demonstrate that the proposed method achieves 91.7% precision, 89.6% recall, and 90.3% mean average precision at 0.5 IoU threshold, outperforming YOLOv8-seg by 7.8% in precision and reducing false positives by 48.4%. Notably, misclassifications of bare rocks as landslides are reduced by 61.9%. The proposed framework effectively suppresses spectral confusion and mitigates class imbalance, delivering robust landslide recognition with precise boundary delineation in complex terrains. This provides an effective tool for automated landslide mapping and geohazard prevention.