High spatial resolution (H $_2$ ) imagery, Houston datasets, interactive learning, semantic gap, spatial–spectral heterogeneity, WHU-Hi datasets, Ocean engineering, TC1501-1800, Geophysics. Cosmic physics, QC801-809

Hyperspectral imagery with high spatial resolution (H2) imagery can synchronously obtain the spectral and spatial features of objects, thus providing richer information. However, the exacerbated spatial–spectral heterogeneity poses new challenges for classification. In this study, an interactive learning framework was proposed to address the current issues in H2 imagery classification. Specifically, we propose a spectral–spatial purification network (S2PNet) to improve classification accuracy. First, a multistage spectral purification module is designed to purify noisy information and mitigate spectral heterogeneity, achieving interaction between spectral optimization and classification. Second, a global–local mutual guide module is utilized to realize image–pixel-level feature interaction, thus enhancing the spatial discriminability of extracted features and reducing spatial heterogeneity. Third, the introduction of dual-stream semantic progressive module facilitates shallow-deep feature interaction, reducing the semantic gap in internal network and enabling a smoother information flow. We validated our approach using the public WHU-Hi hyperspectral datasets and large-scale Houston datasets. Experimental results demonstrate that S2PNet achieves the highest classification accuracy across all tests, significantly outperforming state-of-the-art methods.