Multipath fusion transformer network for hyperspectral remote sensing image classification

To address the challenges of insufficient spectral feature extraction and feature redundancy in hyperspectral image (HSI)classification, this paper proposes a Multi-Path Fusion Transformer Network (MPFT) that integrates Vision Transformer, hybrid residual convolution, and attention mechanisms. The proposed approach effectively utilizes both spatial and spectral information of pixels to overcome these limitations. The MPFT architecture comprises a Data Processing Module (DPM), a Dual-Stream Feature Extraction Module (DSFE), and a Vision Transformer (ViT) module. Specifically, the DPM performs dimensionality reduction and merges HSI information in three-dimensional space, while the DSFE extracts deeper features, which are subsequently processed by the ViT module for classification. Experiments conducted on three benchmark datasets—Indian Pines, University of Pavia, and Salinas—demonstrate that MPFT significantly outperforms mainstream models. Notably, MPFT achieved improvements in Overall Accuracy (OA) by up to 5.52%, Average Accuracy (AA) by up to 6.39%, and Kappa coefficient by up to 6.49% compared to state-of-the-art methods. These results highlight that MPFT not only efficiently extracts spatial-spectral features from HSI but also demonstrates superior effectiveness and robustness in various classification tasks.