SAM-Prior-Guided Spatial–Spectral–Temporal Mamba Framework for Affine Misregistration Hyperspectral Change Detection

Hyperspectral change detection (HSCD) aims to identify land surface changes by analyzing spatial and spectral differences between multi-temporal hyperspectral images. However, most existing methods are designed under the assumption of strict pixel-level correspondence, making them intrinsically sensitive to affine misregistration that frequently occurs in real-world scenarios due to sensor drift, platform instability, or accumulated georeferencing errors. To address this issue, we propose a HSCD framework that integrates image registration and change detection through object-level structural priors. Specifically, the Segment Anything Model (SAM) is employed to extract semantically consistent region masks from pseudo-color projections of hyperspectral images. These masks guide both object-level alignment and sparse label propagation, enabling structure-aware supervision without requiring dense annotations. The expanded labels, encoding object-level priors, are then used to supervise a spatio-spectral-temporal change detection network built upon convolutional and Mamba modules. This architecture effectively models spatial, spectral, and temporal dependencies while maintaining robustness against affine misregistration. Extensive experiments on three HSCD benchmarks demonstrate that our method consistently outperforms existing approaches under both well-aligned and affinely misregistered conditions. The code is available at https://github.com/Jiahuiqu/SAMPGSST- MCDNet.