Hyperspectral image classification via deep network with attention mechanism and multigroup strategy

Jun Wang, Jinyue Sun, Erlei Zhang, Tian Zhang, Kai Yu, Jinye Peng

Abstract: In order to address the correlation between feature maps and the context information of spectral sequences that are not considered in the classi- fication of hyperspectral image (HSI) classification, we proposed a novel deep network with the Attention Mechanism and Multigroup Strategy (AMM-Net), which combines a spatial feature extraction module with attention mechanism and a spectral feature extraction module with multi-grouping strategy for improving classification accuracy. The spatial feature extraction module with attention mechanism is based on the channel spatial attention module (CSAM), which can deal with the output features of the convolution layer differently, focus on more useful classification features and enhance the expression ability of the model. The spectral feature extraction module with multi-grouping strategy includes a long and short time memory (LSTM) network to explore more spectral information and features. Finally, the deep spatial-spectral features are inputs of the softmax classifier. In the experiment, three HSIs benchmarks are utilized to evaluate the performance of the proposed method. The experimental results show that the AMM-Net has excellent performance.