GeoSlide-XMamba: A Spectral-Topographic Boundary-Aware State-Space Network for Landslide Semantic Segmentation

Rapid and reliable landslide mapping from satellite observations is essential for hazard assessment, emergency response, and reservoir-area risk management, yet automatic segmentation remains challenging in mountainous regions because landslide scars are spectrally heterogeneous, terrain-constrained, morphologically irregular, and frequently confused with other exposed surfaces. This study proposes GeoSlide-XMamba, a terrain-conditioned spectral-topographic boundary-aware state-space network for pixel-wise landslide semantic segmentation. The model first separates Sentinel-2 spectral bands and DEM/slope-derived topographic layers into modality-specific branches, integrates them through spectral-topographic adaptive fusion (STAF++), and then performs terrain-conditioned selective state-space scanning in the XMamba bottleneck. Unlike direct token concatenation, the proposed bottleneck uses terrain descriptors to dynamically weight directional selective scan branches so that long-range feature propagation is guided by slope-related morphology. Boundary-aware decoding, signed-distance supervision, and hard-negative mining are further introduced to improve inventory-oriented geometric quality and suppress common false positives. Experiments were conducted on the Landslide4Sense benchmark using 14-channel multispectral-topographic inputs. Among the compared methods, GeoSlide-XMamba achieved the highest validation performance under a unified five-seed protocol, with precision = 0.729, recall = 0.626, F1-score = 0.673, IoU = 0.507, kappa = 0.666, Boundary-F1 = 0.466, and HD95 = 3.45 pixels. Five-seed experiments produced F1 = 0.673 ± 0.003, IoU = 0.507 ± 0.002, Boundary-F1 = 0.466 ± 0.002, and HD95 = 3.45 ± 0.13 pixels, with a 95% CI of [0.670, 0.676] for F1. Relative to the strong 14-channel concatenation baseline, the proposed model improves mean F1 by 0.045 and reduces HD95 by 1.42 pixels. Expanded qualitative inference on Jinsha River patches indicates that the learned spectral-topographic representation transfers plausibly to high-relief reservoir-canyon terrain. These results show that terrain-conditioned state-space modeling can improve both segmentation accuracy and boundary geometry for remote sensing landslide mapping.