High Spectral Image Classification Model Based on Residual-Corrected Adaptive Jigsaw Capsule Network

To improve the precision of hyperspectral image classification, this paper introduces an Adaptive Capsule Residual Jigsaw Network (abbreviated as ACRJ-Net), which builds upon the capsule network structure with an integrated residual correction. Firstly, considering the intricate non-linear inter-band characteristics and global interactions of hyperspectral images, we design a Band Attention Module (BAM) to highlight key features and reinforce inter-feature correlations. Secondly, to thoroughly harness spatial features and optimize local representations, we improve upon the foundational capsule network by integrating a jigsaw-puzzling concept. To ensure its seamless fusion with the capsule network, an adaptive pooling layer is proposed to balance feature dimensions. Lastly, to counteract gradient vanishing in the jigsaw capsule network and enhance feature representation capacity, a residual algorithm module is introduced. This ensures effective fusion and optimization of deep features, ultimately boosting the model’s classification performance. Experimental results reveal that ACRJ-Net surpasses other similar models in terms of overall classification accuracy on the KSC, Pavia University, and Salinas datasets, showing average improvements of 3.48%, 5.80%, and 3.49%, respectively. Ablation experiments in comparison with the base matrix capsule network demonstrate increases in overall accuracy by 1.44%, 1.30%, and 0.33% respectively, validating the model’s efficacy.