Intra- and inter-Mamba for multi-domain hyperspectral image generalization

Abstract Hyperspectral image classification, which aims to assign a unique class label to each pixel, has garnered growing attention in recent years. Although many single-source domain adaptation or domain generalization techniques for hyperspectral image classification have been developed, their generalization capability remains limited. This is primarily because single-source domain adaptation methods are typically tailored to a specific target domain, while single-source domain generalization methods have limited access to diverse source domain variations, restricting their applicability to practical scenarios involving multiple datasets or heterogeneous data sources. To address this issue, we propose a novel intra- and inter-Mamba-based multi-domain generalization network called IMDGN for hyperspectral image classification. First, to overcome the limited receptive field of convolutional neural networks and the high computational complexity of Transformers, an intra- and inter-Mamba network was proposed to capture long-range spectral-spatial dependencies in hyperspectral images. Furthermore, to address variations in the spectral response introduced by heterogeneous sensors, we designed a plug-and-play data augmentation (PDA) module. This module generates multi-view hyperspectral data through amplitude scaling and noise injection, thereby improving robustness to illumination variations and noise perturbations. Experiments on four public benchmarks demonstrate the superiority of the proposed method over existing approaches.