Enhanced Hyperspectral Image Classification Through Dual-Path Channel-Attention Network

The challenge of extracting robust feature representations from hyperspectral images (HSIs) while optimizing model efficiency, particularly with limited training samples, remains a focal issue. This study introduces a novel approach to HSI classification, termed the dual-path channel-attention network. Initially, a bespoke spectral feature extractor is incorporated into the spectral subnetwork to facilitate the learning of spectral correlations. Subsequently, the spatial subnetwork is employed to capture spatial details, integrating spectral data to complement the spectral subnetwork. The two subnetworks are merged through fusion layers, balancing spatial and spectral traits to generate spectral-spatial vectors. Finally, the Softmax regression layer predicts the probability distribution of diverse land features. The method's efficacy is validated using three benchmark hyperspectral datasets: Kennedy Space Center (KSC), Pavia University (PU), and Salinas Valley (SV). With only 10% of training samples, the KSC dataset achieved high accuracy, with overall accuracy (OA), average accuracy (AA), and Kappa coefficient reaching 99.31%, 98.85%, and 99.23%, respectively. Similarly, on the PU dataset, using 5% of training samples, the method achieved OA, AA, and Kappa of 99.65%, 99.57%, and 99.54%, respectively. For the SV dataset, utilizing 5% of training samples, the method yielded OA, AA, and Kappa of 99.81%, 99.79%, and 99.79%, respectively. The findings demonstrate the method's capability to enhance feature extraction from HSIs under constrained sample conditions.