Lightweight Hyperspectral Domain Generalization via Sample-Specific Low-Rank Basis Mixture

Hyperspectral image (HSI) classification often suffers from substantial cross-domain shifts arising from sensor heterogeneity, varying acquisition conditions, and scene-dependent spectral distortions—all of which significantly degrade model robustness in previously unseen environments. Moreover, most existing deep learning approaches rely on heavy architectures, hindering their deployment on resource-constrained platforms such as UAVs or onboard processors. To address these challenges, we propose the adaptive basis spectral network (ABSNet), a lightweight domain generalization (DG) framework that learns adaptive low-rank spectral bases for constructing sample-specific low-rank adapters. By integrating spectral–spatial representations through cost-efficient token modulation, ABSNet is designed to improve robustness to distribution shifts while retaining discriminative spectral information. In addition, the rank-constrained formulation retains representation capacity with minimal parameter overhead, making the framework suitable for edge-level deployment. Experiments on the Houston, Pavia, and HyRANK benchmarks show that ABSNet achieves accuracy comparable to state-of-the-art (SOTA) methods with only a small fraction of the parameters of the compared methods. Ablation studies assess the contribution of each component within the proposed framework. The code will be made publicly available athttps://github.com/qiujialiang/ABSNet