Application of Group-Aware Hierarchical Transformation Based on Grouping in Hyperspectral Image Classification

Currently, deep learning algorithms shine in the field of hyperspectral image (HSI) classification and processing of hyperspectral images. Still, due to the limitations of the input size sensing field of deep learning algorithms, it is generally necessary to cut the input of HSI into fixed-size image chunks first, and then carry out the classification operation. This approach is limited by the size characteristics of HSI. In recent years, the visual transformer (ViT), as a deep learning algorithm that can be applied to multi-tasks, has not surprisingly been introduced to the HSI classification task, which is capable of selecting appropriate patch scales for different input maps through a flexible input feature selection mechanism. Nevertheless, when ViT deals with HSI with many bands, the features extracted by its global self-attention mechanism may exhibit a tendency of over-dispersion, which affects the classification performance. To address the above challenges, this paper proposes a ViT-based packet-aware hierarchical converter (PSHC) architecture that aims to optimize the classification of HSI. The architecture strengthens the local relationships within all spectral channels by embedding a pixel group mapping (PGM) module, which applies a global self-attention mechanism to the algorithm for extracting global spatial-spectral features and achieves better characterization by fusing global and local features. In addition, PSHC adopts a hierarchical structural design, which firstly effectively extracts and fuses multiscale features without enlarging the number of model parameters, significantly improving the classification accuracy. Experimental validation on two benchmark HSI datasets shows that the proposed PSHC method outperforms previous deep learning algorithms in HSI classification tasks, especially in city-like map applications.