HCMA: Hybrid-Convolutional Multi-Attention

Today, hyperspectral images are playing an increasingly important role in the field of remote sensing image analysis. One of the most commonly used methods in the field of image processing today is Convolutional neural network (CNN). HSIs are images containing different wavebands. Using CNN for HSI classification requires CNN to be able to extract both spatial and spectral information of HSI. Most of the previous methods are based on 2-D CNN or 3-D CNN. Few methods have introduced attention mechanisms. For HSI classification tasks, we proposes a model named Hybrid-Convolutional Multi-Attention (HCMA) mechanism. In general, it is a hybrid spectral CNN (HybridSN) combined with Convolutional Block Attention Module (CBAM). The whole model uses 3-D CNN for local spectral information extraction and uses 2-D CNN for local spatial information extraction. In contrast, CBAM performs further extraction of global spatial and global spectral features. This approach combines the advantages of CNN which can perfectly extract the features of local features and attention mechanisms. In order to verify the performance of this method, we conducted strict classification experiments on three typical remote sensing data sets. The proposed model performs better compared to the approaches using only CNN or the attention mechanism.