ASSCA-Net: An Adaptive Spectral–Spatial Cooperative Attention Network for Hyperspectral Image Classification

Abstract

Hyperspectral images (HSIs) possess fine spectral resolution. They can capture continuous and detailed spectral curves of ground objects, providing rich information for accurate classification. However, real-world scenes commonly suffer from diverse ground object morphology, spectral variability, and insufficient spatial–spectral feature interaction. These problems make effective joint feature extraction challenging. To address these issues, this article proposes an adaptive spatial–spectral collaborative attention network (ASSCA-Net). First, the network employs a spectral-guided spatial aggregation module (SSAM) at the input stage to suppress noise and enhance intraclass consistency. Then, a parallel dual-branch architecture is adopted. The spectral branch achieves differentiated band perception through an anchor-driven spectral attention mechanism (ASAM). The spatial branch captures long-range dependencies of heterogeneous targets via dynamic multiscale strip pooling (DMSP). Finally, a bidirectional collaborative attention fusion (BCAF) module enables deep interaction and adaptive fusion of spatial–spectral features. Experimental results on multiple public datasets show that the proposed method outperforms existing mainstream methods in accuracy.