Mapping retrogressive thaw slumps from remote sensing data using morphology-sensitive edge-aware VMamba

Accurate mapping of regional retrogressive thaw slumps (RTSs) provides spatial evidence of the surface expression of thaw-related mass wasting and can support subsequent process and ecological-impact assessments. However, existing RTS mapping methods often suffer from limited adaptability to RTSs with varying sizes and morphologies and blurred boundary delineation, making it difficult to capture the characteristics of RTSs with pronounced spatial heterogeneity and composite slump complexes. Therefore, this study proposes Morphology-Sensitive Edge-Aware VMamba (MSE-VMamba) for reliable RTS mapping. First, the model employs a VMamba backbone with a hierarchical architecture and global dependency modelling capability to support consistent representation of RTSs across varying object sizes, shapes, and spatial configurations. Afterwards, an explicit edge-aware mechanism combined with the gray level co-occurrence matrix–derived (GLCM-derived) texture features is integrated into the architecture, where frequency- and spatial-domain information jointly increase the sensitivity to boundary transitions and surface structural differences, enabling more precise edge delineation. The validation results for the Tibetan Plateau demonstrate that MSE-VMamba achieves a pIoU of 0.7026, representing a 5.54-percentage-point improvement over the strongest tested CNN-based baseline. Furthermore, the model performs robustly across RTS with varying sizes and morphologies, effectively reducing misclassification between adjacent slumps.