Abstract
Hyperspectral image (HSI) classification is a fundamental remote sensing task, but the high-dimensional nature of HSI data poses challenges for acquisition, storage, transmission, and processing. To mitigate these issues, researchers have developed compressive classification methods based on the coded aperture snapshot spectral imaging (CASSI) system, enabling direct classification in the compressed domain without reconstructing the full hyperspectral cube. However, existing methods often fail to fully leverage spatial-spectral correlations, suffer from limited multi-scale feature extraction, and are constrained by fixed receptive fields and degradation caused by compression. To overcome these limitations, we propose ECSA-UNet, an end-to-end framework that directly classifies compressive spectral measurements from CASSI. The model introduces a spatial-spectral decoupling module and refines the U-Net architecture with global–local feature extraction, Spatial-Spectral Attention, a Multi-Scale Enhancement Pyramid and Cross-Stage Connections, enabling dynamic receptive fields, improved long-range dependency modeling, and robust local–global feature fusion. Moreover, the CASSI coded aperture is parameterized as a trainable component, allowing for joint optimization of coding and classification within an end-to-end learning paradigm. Extensive experiments on four benchmarks show ECSA-UNet achieves state-of-the-art accuracy, and ablations confirm each module’s contribution. In addition, it maintains an effective trade-off between parameter overhead and computational efficiency, offering a viable solution for latency-sensitive hyperspectral remote sensing tasks.