EAS $$^2$$ KAM: enhanced adaptive source-selection kernel with attention mechanism for hyperspectral image classification

Abstract

Hyperspectral Images (HSIs) possess extensive applications in remote sensing, especially material discrimination and earth observation monitoring. However, constraints in spatial resolution increase sensitivity to spectral noise, limiting the ability to adjust Receptive Fields (RFs). Convolutional Neural Networks (CNNs) with fixed RFs are a common choice for HSI classification tasks. However, their potential in leveraging the appropriate RF remains under-exploited, thus affecting feature discriminative capabilities. This study introduces an Enhanced Adaptive Source-Selection Kernel with Attention Mechanism (EAS\(^2\)KAM) for HSI Classification. The model incorporates a Three Dimensional Enhanced Function Mixture (3D-EFM) with a distinct RF for local low-rank contextual exploitation. Furthermore, it incorporates diverse global RF branches enriched with spectral attention and an additional spectral-spatial mixing branch to adjust RFs, enhancing multiscale feature discrimination. The 3D-EFM is integrated with a 3D Residual Network (3D ResNet) that includes a Channel-Pixel Attention Module (CPAM) in each segment, improving spectral-spatial feature utilization. Comprehensive experiments on four benchmark datasets show marked advancements, including a maximum rise of 0.67% in Overall Accuracy (OA), 0.87% in Average Accuracy (AA), and 1.33% in the Kappa Coefficient (\(\kappa \)), outperforming the top two HSI classifiers from a list of eleven state-of-the-art deep learning models. A detailed ablation study evaluates model complexity and runtime, confirming the superior performance of the proposed model.