Hyperspectral Image Classification With Spatial–Spectral–Channel 3-D Attention and Channel Attention

For the advantages of focusing on important information and avoiding gradient vanishing, the attention mechanism (AM) and the densely connected network (DenseNet) are widely used in the hyperspectral image (HSI) classification. However, AMs applied to HSI classification (HSIC) do not consider different roles of each pixel, and the DenseNets ignore the different importance of different channel features, which affect HSIC performance. To solve the problems above, a novel HSIC method based on the spatial-spectral–channel 3-D AM (3DAM-SSC) and DenseNet improved with channel AM (CAM) is proposed, which is abbreviated as 3DAM-CAM DenseNet. First, the 3DAM-SSC is proposed and applied, the weight of each pixel point is determined by the pixel value itself and the correlations of different pixels in three dimensions. Second, the two-branch network structure with different scale convolutional kernels is used to fuse features. Third, a CAM is added after each DenseNet transition layer, and a total of three CAMs are used. Comparative experiments are conducted on hyperspectral datasets, which are Indian Pines (IP), KSC, and WHU-Hi-HongHu. The overall accuracy (OA) of 3DAM-CAM DenseNet in each dataset is 98.65%, 99.64%, and 98.73%, respectively, which are better than the compared methods.