Congratulations to Rui Qin on His Paper's Acceptance by the Journal of Shanghai Jiao Tong University (English Edition).

Hyperspectral Image Classification Method Based on Global Space-spectral Attention Mechanism

abstract：In hyperspectral remote sensing images, the interactions between pixels within a defined spatial range lead to the mixing of adjacent pixels. Additionally, the high similarity between adjacent spectra results in information redundancy, which hinders the extraction of global spatial and spectral correlations. To address the issues of adjacent pixel mixing and spectral redundancy, this paper proposes a hyperspectral image classification method based on a global spatial-spectral attention mechanism.

First, the global spatial attention module in the proposed method uses multi-scale dilated convolutions to obtain a larger receptive field, capturing global spatial correlations and extracting unmixed pixel information. Then, the global spectral attention module designs a spectral domain partitioning algorithm, which uses the product of local density and information entropy as a threshold to divide the spectrum into scattered subsets, eliminating redundant information. This approach fully utilizes the global contextual information of the entire spectral band and extracts the correlation of global spectral information.

Finally, these two modules are combined to obtain the global correlations of both space and spectrum. Experimental results show that the proposed method achieves overall accuracies of 97.28%, 94.73%, and 95.76% on three WHU-Hi hyperspectral datasets, outperforming the comparison methods.