An End-to-End Automatic Classification Algorithm for Hyperspectral Images via 3D Convolution and Swin-Transformer

Deep Learning (DL) is one of the key technologies to realize artificial intelligence (AI), which achieves automatic classification and prediction of data by simulating the learning process of the human brain. Convolutional neural network (CNN) is a very important architecture in DL, which is being widely used in the classification task of hyperspectral images (HSI).The local convolutional operation in CNN cannot learn the global semantic information in HSI adequately, but the multi-attention mechanism in the Swin-Transformer can fully utilize the rich discriminative information. So we jointly design an "end-to-end" deep network model with Swin-Transformer to further improve the classification accuracy of HSI. First, the HSI is downscaled using principal component analysis, and the neighborhood data around the pixel is selected as the input samples to make full use of the null-spectrum joint information in the image; then, the input samples are converted into sequence feature vectors using convolutional layers; finally, the classification is carried out by using the constructed Swin-Transformer blocks. The experiments show that our method can achieve better classification performance than some existing CNN models.