Spatial-Spectral Transformer for Hyperspectral Image Classification

Recently, a great many deep convolutional neural network (CNN)-based methods have been proposed for hyperspectral image (HSI) classification. Although the CNN-based advantages of spatial feature extraction, they are difficult to handle sequential data with and CNNs not good at modeling long-range dependencies. However, spectra HSI kind data, usually contains hundreds bands. Therefore, it is processing well. On other hand, Transformer model, which based on an attention mechanism, has proved its in data. To address issue capturing relationships long distance, this study, investigated Specifically, new classification framework titled spatial-spectral (SST) In SST, well-designed CNN used extract features, modified (a dense connection, i.e., DenseTransformer) capture relationships, multilayer perceptron finish final task. Furthermore, dynamic augmentation, aims alleviate overfitting problem therefore generalize model well, added SST (SST-FA). addition, limited training samples classification, transfer learning combined another transferring-SST (T-SST) proposed. At last, mitigate improve accuracy, label smoothing introduced T-SST-based (T-SST-L). The SST-FA, T-SST, T-SST-L tested three widely datasets. obtained results reveal that models provide competitive compared state-of-the-art methods, shows concept opens window