Work

Multiscale Graph Transformer Network With Dynamic Superpixel Pyramid for Hyperspectral Image Classification

Year: 2025

Type: article

Abstract: Hyperspectral image (HSI) classification plays a crucial role in remote sensing applications, leveraging the rich spectral and spatial information inherent in HSI. However, the diverse feature scales of ground objects pose challenges for traditional methods, which often employ fixed-scale feature extraction and struggle to adapt dynamically to varying spatial structures. Moreover, conventional approaches typically emphasize either local spectral or spatial features while neglecting global contextual relationships. To address these limitations, we propose a multi-scale graph Transformer network (MSGTN), which captures spatial features at different scales through multi-scale graph convolutional networks with adaptive graph structures. First, we employ differentiable superpixel segmentation to construct flexible graph topologies in the feature space, enabling adaptive representation learning at various spatial scales. Second, a multi-scale graph convolutional network (GCN) is introduced to jointly model spectral and spatial features, enhancing the network's capacity to represent local structural information. Finally, a spectral–spatial Transformer module is integrated to model complex global dependencies among superpixels, capturing long-range contextual relationships. Extensive experiments conducted on four widely used hyperspectral datasets, University of Pavia (UP), Indian Pines (IP), Salinas Valley (SA), and WHU-Hi-HongHu (WHHH), demonstrate the effectiveness of the proposed MSGTN. The model achieves classification accuracies of 99.01% (UP), 96.94% (IP), 99.43% (SA), and 97.95% (WHHH) using only 1%, 5%, 1%, and 1% of labeled training samples, respectively. Compared with state-of-the-art methods, MSGTN exhibits superior classification performance and enhanced robustness across varying data conditions. The code of this work is available athttps://github.com/w-tingting/MSGTN.

Source: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

DOI: 10.1109/jstars.2025.3570267

Authors: Tingting Wang, Yao Sun, Yunfeng Hu

Institution: Jilin University

Language: English

Topic: Remote-Sensing Image Classification

Subfield: Media Technology

Field: Engineering

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $1,800

Funder: China Postdoctoral Science Foundation