Advancing hyperspectral image classification through multihead dual attention mechanisms

Hyperspectral Image (HSI) technology, known for its high spectral resolution, is a crucial remote sensing tool applied in fields like agriculture, environmental monitoring, and urban planning. Effective HSI classification is critical for extracting valuable insights in these applications. However, efficiently capturing key spectral and spatial features, which are vital for distinguishing similar land cover types, remains a significant challenge in current research. To address this challenge and improve classification accuracy, this paper introduces an innovative hyperspectral image classification network: the SpectraSpatial Hyper-Attention Network (SSHANet). SSHANet integrates a Spectral-Spatial Feature Extraction Branch and a Global Spectral Analysis Branch. This allows it to effectively capture local spatial features and interdependencies between spectral channels, enabling precise handling of global and local spectral-spatial correlations in HSI images. SSHANet is designed to maximize the use of complex HSI images through advanced algorithmic optimization, with the goal of enhancing land cover classification accuracy. Experimental results demonstrate that SSHANet achieves comparable or superior classification performance to current state-of-the-art methods, validating its effectiveness.