Attention-enhanced multi-scale dynamic convolution combined pyramid pooling Transformer for small sample hyperspectral image classification

Hyperspectral images contain a greater number of spectral bands, allowing for the accurate capture of complex features through higher spectral resolution compared to RGB images. In hyperspectral image (HSI) classification, each pixel is categorized based on the type of land cover it represents. In recent years, the Transformer framework has shown great potential in the field of hyperspectral image (HSI) classification due to its stronger global modeling capabilities compared to convolutional neural networks (CNN). At the same time, although the traditional CNN model has a strong ability to capture local information, in order to enhance the model to extract more discriminative hyperspectral image features, it is necessary to stack the convolutional layer or increase the number of channels in the convolutional layer, which will lead to computationally expensive. In order to overcome the above problems and take advantage of the advantages of CNN and Transformer, this paper proposes the attention-enhanced multi-scale dynamic convolution combined pyramid pooling Transformer method (AMDCPPT). First, a multi-scale CNN feature extractor is constructed using dynamic convolution. The model capability can be improved without stacking a large number of convolution blocks or increasing the channel dimension. And multi-scale design can also capture multi-scale feature information, making the feature information richer. Secondly, the attention enhancement module processes the features extracted by the CNN composed of dynamic convolution, enhances the useful information for hyperspectral image classification, and suppresses the useless information. Finally, the pyramid pooling Transformer uses its powerful upper global information extraction ability to extract features for classification. In this paper, three standard datasets were used, and experimental analysis confirmed that the classification performance of the proposed method is superior to several methods.