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

Area of Science:

• Earth and Environmental Sciences
• Remote Sensing
• Geomorphology

Background:

• Automatic landslide mapping from satellite data is crucial but challenging due to spectral heterogeneity and irregular shapes of landslide scars in mountainous terrain.
• Existing methods struggle with accurately segmenting landslides, leading to potential inaccuracies in hazard assessment and risk management.
• The spectral and topographic characteristics of landslides are complex and often confused with other surface features, necessitating advanced modeling approaches.

Purpose of the Study:

• To propose GeoSlide-XMamba, a novel terrain-conditioned spectral-topographic boundary-aware state-space network for precise pixel-wise landslide semantic segmentation.
• To improve the accuracy and geometric quality of landslide inventories derived from remote sensing data.
• To develop a model that effectively integrates multispectral and topographic information for robust landslide detection.

Main Methods:

• Developed GeoSlide-XMamba, a state-space network incorporating modality-specific branches for Sentinel-2 spectral bands and DEM/slope-derived topographic layers.
• Implemented spectral-topographic adaptive fusion (STAF++) for integrating diverse data modalities.
• Introduced terrain-conditioned selective state-space scanning in the XMamba bottleneck, guided by slope-related morphology for feature propagation.
• Employed boundary-aware decoding, signed-distance supervision, and hard-negative mining to enhance geometric accuracy and reduce false positives.

Main Results:

• GeoSlide-XMamba achieved state-of-the-art performance on the Landslide4Sense benchmark, outperforming existing methods in landslide semantic segmentation.
• The model demonstrated superior precision (0.729), recall (0.626), F1-score (0.673), and IoU (0.507) under a unified five-seed protocol.
• Significantly improved mean F1-score by 0.045 and reduced HD95 by 1.42 pixels compared to a strong baseline.
• Qualitative inference showed plausible transfer of learned representations to high-relief reservoir-canyon terrain.

Conclusions:

• Terrain-conditioned state-space modeling offers a powerful approach to enhance both segmentation accuracy and boundary geometry in remote sensing landslide mapping.
• The proposed GeoSlide-XMamba effectively leverages spectral and topographic data for improved landslide detection and hazard assessment.
• This methodology holds significant potential for advancing landslide inventory generation and risk management strategies.