Deep Spatial-Spectral Multi-Resolution Attention Residual Network for Hyperspectral Image Classification

Due to its modular design and robust feature extraction capabilities, deep learning has emerged as one of the essential tools in hyperspectral image classification. However, effectively extracting deeper-level features and improving the ability to analyze spatial-spectral information simultaneously remain challenging problems. In response to these issues, a deep spatial-spectral multi-resolution attention residual network is proposed, comprising two critical components: a multi-level transfer fusion residual network and a spatial-spectral multi-resolution fusion attention residual network. By employing different residual blocks, the multi-level transfer fusion residual network can effectively promote interactions between feature information and obtain deeper features. Furthermore, the spatial-spectral multi-resolution fusion attention residual network can adaptively extract and integrate spatial-spectral joint features and multi-resolution features comprehensively from hyperspectral data. The performance was evaluated against other methods on the Pavia University hyperspectral dataset. The experimental results demonstrate that the network proposed in this paper exhibits promising classification performance.