An unsupervised spatial–spectral learning framework for high-resolution hyperspectral landslide extraction

Landslides often result in severe casualties and substantial economic losses, and timely large-scale landslide extraction is crucial. However, the advancement of automated landslide extraction remains constrained by limited labeled data and uneven regional coverage. Therefore, developing unsupervised landslide extraction methods that efficiently identify new landslide instances and dynamically update inventories is of great importance. Existing unsupervised approaches based on multispectral data often suffer from misclassification due to the spectral similarity among different surface materials. Although high-resolution hyperspectral imagery provides enriched spectral and spatial information, systematic unsupervised spatial–spectral learning strategies remain underexplored. To address these challenges, this study proposes an unsupervised spatial–spectral learning framework. Specifically, a spectral-derivative-enhanced hierarchical superpixel modeling is employed to capture fine-scale spatial representations. A Mamba module enhanced with prototype-guided spectral attention is proposed for hyperspectral spectral feature extraction. These spatial and spectral representations are further integrated through cross-patch contrastive graph learning to achieve structure-consistent spatial–spectral fusion. The proposed unsupervised approach achieved an Intersection over Union (IoU) of 76.56% and an F1 score of 86.72% in the study area. The proposed methodology establishes a scalable and annotation-free paradigm for hyperspectral landslide extraction, offering a theoretically grounded solution for large-area and continuously updated landslide inventory generation.