DAFNet: A Dual-Branch Attention Fusion Network with Multi-Scale Residual Learning for Hyperspectral Image Classification
Abstract
Accurate classification of hyperspectral images (HSI) hinges on effective fusion of spectral and spatial information while mitigating challenges of high dimensionality, redundancy and limited labeled samples. We present DAFNet, a dual-branch fusion network that integrates multi-scale residual blocks (MSR-Block) with spectral- and spatial-wise attention modules (SAE-Module and PAF-Module). The MSR-Block simultaneously applies <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1 \times 1,3 \times 3,5 \times 5$</tex> and dilated convolutions to capture fine- to coarse-scale context. SAE-Module adaptively reweights spectral channels, and PAF-Module highlights salient spatial regions. A learnable fusion layer balances spectral and spatial cues, producing highly discriminative feature maps. Evaluated on four benchmark datasets (Indian Pines, WHU-Hi-LongKou, Salinas and Xuzhou), DAFNet consistently outperforms five state-of-the-art methods, achieving up to 2.5% absolute gains in Overall Accuracy and substantial improvements in Average Accuracy and Kappa coefficient. Qualitative analyses confirm sharper class boundaries and reduced misclassifications in mixed - landcover areas. DAFNet offers a robust, end-to-end solution for HSI classification, with potential for real-world remote sensing and precision agriculture applications.