MSSTFormer: Multiscale Spectral–Spatial Supertoken Aggregation Transformer for Hyperspectral Image Classification

Le résumé fourni par la source

Transformer-based methods have attracted attention in hyperspectral image (HSI) classification due to their powerful global modeling capability. However, the existing Transformer-based methods lack mechanisms to reinforce token representations, which leads to redundant features and limited discrimination ability. In addition, local spatial variations and intra-group spectral differences are prevalent in HSIs across different scales. It is difficult for the existing Transformer-based methods to capture both local spatial information and global spectral changes, which limits the performance of such methods on complex spectral-spatial coupled datasets. To address these problems, a novel multi-scale spectral-spatial supertoken aggregation transformer (MSSTFormer) is proposed to achieve effective fusion of multi-scale spectral-spatial information. Specifically, MSSTFormer introduces a novel token representation (i.e., Supertoken) that aggregates spectral-spatial local features into a compact token sequence. Then, the modeling of the spectral-spatial correlation among tokens is enhanced using a chained dual-attention mechanism that incorporates the geographical neighborhood prior. Finally, a simple classification head with adaptive weight allocation integrates multiscale features for pixel-wise classification. We compare the proposed method with the SOTA methods (traditional methods, convolutional neural network (CNN)-based methods, transformer-based methods) on four HSI datasets, and the experimental results show that the proposed MSSTFormer method has higher classification accuracy.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.