Edge-Aware Fusion Network for Accurate Hyperspectral Image Classification

Hyperspectral images (HSIs) are gaining traction in consumer electronics for applications such as mobile health diagnostics, environmental monitoring, and agricultural quality assessment. However, effective practical deployment requires a rigorous balance between precise spectral-spatial feature extraction and computational efficiency. Existing hyperspectral image classification(HSIC) methods often fail to preserve oriented textures and edge details, limiting their usability in consumergrade systems. To address these challenges, we propose the Edge-Aware Fusion Network (EAF-Net), a multi-branch architecture that enhances edge preservation and robust spectral-spatial feature fusion. EAF-Net integrates: (1) An Oriented Feature Emphasis Spatial Branch that employs an Oriented Feature Emphasis Module to extract orientation-sensitive texture and boundary features. (2) A Multi-Scale Attention Spectral-Spatial Branch that groups spectral bands and applies spectral as well as multi-group spatial attention to emphasize informative bands and spatial contexts. (3) A Parallel Convolution Spectral Branch with compact 1×1 convolutions and non-linear convolution kernels to model cross-band interactions with low overhead. Experimental results on four public hyperspectral datasets show that, under the same training/testing protocol, EAF-Net outperforms selected baselines in terms of OA, AA, and Kappa, with notable gains in boundary pixel classification. The network, implemented with GPU acceleration, demonstrates strong performance and potential as a high-accuracy framework for hyperspectral image analysis.