3D Lightweight Spatial-Spectral Attention Network for Hyperspectral Image Classification

Compared with the convolutional neural network (CNN), a 3D lightweight network (3D-LWNet) can successfully perform hyperspectral image (HSI) classification using fewer network parameters. However, results of these methods can not reflect the effect of each of the different features used in the classification process. To address this problem, we propose a 3D lightweight spatial-spectral attention network (3D-LSSAN), which adopts a 3D large kernel attention (3D-LKA) mechanism to better determine the effect of each feature in HSI classification. 3D-LKA directly operates on spatial-spectral features of HSI, which can reduce the number of trainable parameter and overcome the network overfitting. Moreover, 3D-LKA disassembles the convolution kernel, which can capture local information and long-range information in HSI using a small amount of computational power. Specifically, the HSI is first sent through a convolutional network layer to generate shallow spatial-spectral features. Second, these shallow features are input into six spatial-spectral attention (SSA) units based on 3D-LKA to emphasize the importance of different parts of features. Finally, output features of the SSA units are fed into a fully connected layer to obtain the classification result. Experimental results on two publicly available data sets demonstrate that the proposed 3D-LSSAN achieves better classification performance than the other techniques.