Central spectral-spatial context attention for hyperspectral image classification

Hyperspectral image classification assigns a class label to each pixel in images with hundreds of spectral bands, and recent deep learning methods have achieved strong performance. However, many existing approaches suffer from high computational complexity, leading to long training and inference times as well as heavy memory consumption. Moreover, they often extract redundant spectral information while insufficiently exploiting local spatial context, which limits efficiency and generalization. To address these challenges, we propose an efficient central spectral–spatial context attention framework that emphasizes informative spectral characteristics while incorporating discriminative local spatial information. The framework introduces a Center Sub-Cube Attention (CeSCA) block embedded within residual blocks built using separable convolutions. Motivated by the higher discriminative power of central spectral features, the CeSCA block derives attention from a central sub-cube to jointly capture spectral and local spatial cues. This attention mechanism recalibrates the full three-dimensional feature map, enhancing discriminative spectral–spatial representations while reducing redundancy and computational cost. Extensive experiments on the Indian Pines, Pavia University, and Salinas Scene datasets demonstrate that the proposed method achieves state-of-the-art accuracy under both 10% and 30% training sample ratios. Computational cost analysis further confirms clear advantages in runtime, parameter count, and computational complexity.