BACA-Net: a band-adaptive collaborative attention network for hyperspectral image classification

This paper proposes BACA-Net, a Band-Adaptive Collaborative Attention Network for hyperspectral image classification. The proposed method addresses the challenge of effectively extracting discriminative spectral-spatial features through a novel collaborative attention framework. BACA-Net integrates three parallel attention modules – global, local, and channel attention – with an interaction matrix to enable adaptive learning of significant spectral bands and spatial features. The architecture implements a dynamic feedback mechanism that enhances feature representation through multi-level fusion and adaptive feature optimisation. Additionally, the method introduces a multi-dimensional statistical band representation approach that comprehensively characterises spectral properties by integrating mean, root mean square, and maximum value features, capturing subtle spectral differences effectively. Experimental validation on four benchmark datasets (Indian Pines, Pavia University, Salinas Valley, and Houston2013) demonstrates that BACA-Net significantly outperforms existing CNN-based and attention-based methods, achieving overall accuracies of 93.21%, 96.47%, 96.54%, and 93.13%, respectively. The method exhibits exceptional performance in classifying spectrally similar categories and those with limited training samples. Computational efficiency analysis shows that BACA-Net maintains low inference times (0.228–0.887 ms) and moderate GPU memory consumption while delivering superior classification performance, making it a promising solution for practical hyperspectral remote sensing applications.