Supervised Spatially Spectrally Coherent Local Linear Embedding—Wasserstein Graph Convolutional Network for Hyperspectral Image Classification

Hyperspectral Image (HSI) classification faces challenges due to the high-dimensional and complex nature of the data. Graph Convolutional Network (GCN) is a widely used method for Hyperspectral Image (HSI) classification. Conventional GCN analyzes fixed adjacency relations and ignores spatial spectral relationships. To overcome these issues, we propose a Supervised Spatially Spectrally Coherent Local Linear Embedding integrated with Wasserstein Disatance (WD) based Graph Convolutional Network (S<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>CLLE-WGCN) for HSI classification. This method combines the strengths of Local Linear Embedding (LLE) for dimensionality reduction and Wasserstein Distance (WD) for graph adjacency matrix construction, named Wasserstein Graph Convolutional Network (WGCN) for effective feature extraction in non-Euclidean spaces. The S<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>CLLE branch enhances the spectral-spatial manifold representation preserving local geometric variations and class boundaries, while effectively reducing the dimensionality of the dataset, while the Earth mover’s or Wasserstein Distance better discriminates hyperspectral classes. The S<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>CLLE-WGCN achieves state-of-the-art average accuracies: 98.44% (Indian Pines), 99.35% (Houston), 97.92% (Pavia), and 99.74% (Salinas Scene). These results demonstrate that S3CLLE-WGCN outperforms the baselines in terms of various evaluation criteria.