SD-DSSAN: synergistic denoising and Dual-Path sparse shuffling attention network for hyperspectral image classification

Hyperspectral image (HSI) classification is a demanding undertaking, as the information redundancy and inherent noise interference of the data create barriers to effective feature acquisition. Many existing methods struggle to simultaneously suppress noise interference, reduce information redundancy, and capture discriminative information features, while also facing difficulties in effectively modeling multi-scale geometric structures and cross-scale contextual dependencies, which constrains their efficacy in managing sophisticated HSI classification objectives. To tackle the aforementioned challenges, we present a synergistic denoising and dual-path sparse shuffling attention network (SD-DSSAN). Three novel synergistic components are integrated within the proposed architecture: 1) a hybrid convolutional additive denoising (HCAD) module is proposed, which suppresses noise interference and provides more reliable spectral-spatial representations for downstream processing; 2) a multi-branch directional convolution (MBDC) module is proposed, which leverages directional kernels across a multi-path architecture to capture diverse geometric structures and mitigate redundant information; and 3) a dual-path sparse shuffling attention (DSSA) mechanism is proposed, which combines sparse sampling with channel shuffling to enable cross-scale information interaction and extract highly discriminative information features. Extensive experiments on four benchmark datasets (Pavia University, Salinas, WHU-Hi-HanChuan, and WHU-Hi-HongHu) demonstrate that SD-DSSAN achieves superior performance compared with representative methods.