Hyperspectral Image Classification Based on Spatial Spectral Attention and Pre-Activation Residual Networks

‍ ‍In many deep learning algorithms for hyperspectral image classification， the classification performance of their models needed to be improved due to the poor discriminant representation of extracted spatial spectral features. To solve this problem， this paper proposed a hyperspectral image classification algorithm based on spatial spectral attention and pre-activation residual network. Firstly， a spatial spectral feature extraction module based on the spatial spectral attention mechanism was designed to recalibrate the spatial spectral features， so as to ensure that the spatial spectral features can focus on more discriminative channels and spatial positions during subsequent joint learning； secondly， a joint learning module of spatial spectral features based on the pre-activation residual network was designed， in which the pre-activated residual network improved the network structure of the original residual building block， so that it could capture more discriminative deep spatial spectral features during joint learning of spatial spectral features recalibrated by attention mechanism， and therefore the classification performance of the classifier could be improved. The experimental results showed that compared with some existing hyperspectral image classification algorithms， the proposed algorithm could attain higher classification accuracy. It indicated that the algorithm can effectively obtain spatial spectral features representation with stronger discriminative ability.