I2SNet: A Lightweight Coarse-to-Fine Hierarchical Network for Hyperspectral Image Classification From Image to Subpixel

Abstract

Hyperspectral image classification (HSIC) faces critical challenges, including restricted receptive fields in patch-level methods and coarse feature representations alongside subpixel blindness in conventional image-level networks. To address these bottlenecks, we propose I2SNet, a novel lightweight image-level end-to-end architecture driven by a coarse-to-fine hierarchical feature learning paradigm. The network systematically extracts spatial–spectral representations across patch, pixel, and subpixel scales. Specifically, a spatial–spectral linear transformer (SSLT) and a spatial–spectral deformable Mamba (SSDM) are unified to efficiently capture global patch interactions with linear complexity. Subsequently, multiscale spatial–spectral dynamic convolutions (MS3DCs) extract fine-grained local contexts. Furthermore, a subpixel nonlinear mixing model (SNLMM) is introduced to explicitly address subpixel blindness by modeling the physical mixing mechanisms of diverse materials. Extensive experiments on four benchmark datasets demonstrate that I2SNet achieves strong classification accuracy while maintaining an efficient and compact computational profile compared to state-of-the-art approaches. To facilitate reproducibility, our code is publicly available at https://github.com/flyzzie/I2SNet