Multi-view graph representation learning for hyperspectral image classification with spectral–spatial graph neural networks
Résumé
Hyperspectral image (HSI) classification benefits from effectively handling both spectral and spatial features. However, deep learning (DL) models, like graph convolutional networks (GCN), face challenges in computation time, overfitting, and less informative features. To address these challenges, we propose a novel method called MV-GRL (multi-view graph representation learning) for HSI classification using spectral–spatial graph neural networks. MV-GRL incorporates spectral and spatial features with the extended morphological profile (EMP) and employs a multi-view graph autoencoder (MVGAE) to learn a low-dimensional latent graph representation by fusing two input graphs: spectral features-based graph and spatial-based graph. This fusion enhances the discriminative power of the features. Additionally, we introduce a semi-supervised spectral–spatial GCN using the multi-view latent representation, leveraging labeled and unlabeled samples for improved classification performance. By leveraging both labeled and unlabeled data, our method effectively captures underlying relationships and enhances overall accuracy. Experimental results on the Indian Pines, Salinas, and Pavia University datasets demonstrate its competitive performance, achieving overall accuracy (OA) scores of 96.91%, 97.64%, and 98.88%, respectively, surpassing state-of-the-art (SOTA) models.