Prototypical Spectral Basis Learning for Effective Cross-Scene Few-Shot Hyperspectral Image Classification

Existing few-shot hyperspectral image (HSI) classification methods predominantly concentrate on either mining the sparse semantic knowledge only in the target HSI with limited pixel-wise annotations or further transferring the semantic information from a source domain with the same categories. However, the semantic knowledge that can be transferred across scenes (i.e., images with diverse categories) is scarcely taken into account for performance enhancement. To mitigate this problem, we first formulate the cross-scene few-shot HSI classification (CSFS-HSIC) task and then propose a novel prototypical spectral basis learning-based CSFS-HSIC method. Specifically, we develop a prototypical spectral basis learning network (PSBLN) to extract spectral bases from an auxiliary HSI dataset with extensive pixel-wise annotations. These bases are not only interdiscriminative but also capable of accurately linearly reconstructing the prototypical spectra of all categories, enabling the network to distill cross-scene transferable semantic knowledge in the spectral domain. Moreover, we present a knowledge-enhanced HSI classification network comprising several novel knowledge-calibrated cross-Mamba (KCCM) modules. These modules can seamlessly inject the cross-scene transferable semantic knowledge and the long-range spatial dependency into feature representation, thus producing better generalization capacity in unseen target HSI even with scarce pixel-wise annotations. Sufficient experimental results verify the efficacy of the proposed method in terms of CSFS-HSIC. The code will be available at https://github.com/zhaoxb2025/PSBL-main