Multi-Scale Semantic Selection and Spatial Constraint-Guided Network for Cross-Scene Hyperspectral Image Classification

Cross-scene classification of hyperspectral images attracts extensive research attention due to the prominent distribution discrepancies existing in hyperspectral images across different scenes. Most existing methods mitigate domain shift by expanding source domain samples and aligning feature distributions. However, these approaches fail to fully explore the spatial semantics of samples during the expansion process. Moreover, features are compressed into one-dimensional vectors in the alignment stage, resulting in the loss of critical spatial location information. To address the above issues, this paper proposes a multi-scale semantic selection and spatial constraint-guided network (MSCGnet). Specifically, the multi-scale semantic selection generator adopts a spatial diffusion scanning strategy to optimize the token serialization rule of Mamba. Pixels are arranged from the center to the periphery to maintain spatial continuity. Combined with the multi-scale semantic selection routing, multi-scale spectral–spatial features are extracted and a semantic selection matrix is constructed to guide Mamba to generate diverse augmented samples. The spatial constraint-guided discriminator leverages class activation map projection to impose explicit spatial constraints on feature distributions, further improving the reliability of augmented samples. Comprehensive experiments on multiple cross-scene HSI datasets demonstrate that the proposed method achieves superior classification accuracy and generalization performance with low model complexity.