RA-ViT:Patch-wise Radially-Accumulate Module for ViT in Hyperspectral Image Classification

Hyperspectral Images(HSIs) are data containing abundant spatial and spectral information, which is collected by advanced remote sensors. HSI Classification is a pixel-wise classification task that has broad prospects in the era of science and technology. In recent years, the widely used convolutional neural networks (CNNs) have come to the leading place in HSI Classification. However, the lack of utilization of spatial information limits its further application. To solve this issue, we considered the recently proposed Vision Transformer(ViT), which is modularized structures that are entirely based on self-attention mechanism. Furthermore, we proposed a patch-wise radially-accumulate module for ViT(RA-ViT) in HSI Classification. We evaluated the proposed method on Indian Pines(IP) and Kennedy Space Center(KSC) datasets. The results of experiments demonstrate the effectiveness of RA-ViT with comparison to current advanced models.