S2M2-GFA: spectral–spatial multistream Mamba-driven granularity-fusion alignment for cross-domain few-shot image classification

Hyperspectral remote sensing of Earth is a cornerstone for dynamic environmental monitoring and fine-scale management, yet its deployment across scenes suffers from “domain shift”: the source domain is label-rich while the target domain is almost label-free, causing classification accuracy to plummet. This has become a critical bottleneck limiting the efficient use of environmental remote-sensing data. Existing methods mostly rely on forced feature alignment, which tends to distort the spectral-spatial structure. This paper leverages the linear long-range modeling power of Mamba to “skip” forced alignment and proposes the Spatial-Spectral multistream Mamba-driven Granularity-Fusion Alignment Network (S2M2-GFA). Instead of performing explicit mapping in the feature space, the method directly bridges inter-domain distributional gaps via global sequence modeling with linear complexity. This study designs a Spatial-Spectral Multi-stream Extraction and Fusion module (S2MSEF). Multi-path parallel convolutions couple global spectra with local textures, pixel-band bidirectional interaction pre-suppresses intra-class spectral variance and inter-class spectral confusion, and compact robust joint features are output. And the Mamba-driven Granularity Fusion Alignment module (MDGFA) is proposed. Forward-backward dual-path Mamba injects fine texture, mid-level structure, and broad semantics layer-wise into a shared multigranularity pixel space, enhancing inter-class discrimination. S2M2-GFA delivers a ready-to-deploy AI solution for accurate cross-scene hyperspectral monitoring of the environment under scarce labels.