Graph Kolmogorov–Arnold Convolutional Network for Hyperspectral Image Classification

The Graph Neural Network (GNN) exhibits noteworthy performance in hyperspectral image classification (HSIC) due to its efficient message passing structure, which employs the Multi-Layer Perceptron (MLP) for feature extraction and utilizes nonlinear activation function between two layers of the network. However, two significant limitations (poor model scaling and lack of interpretability) impede GNN's ability to extract deeper features in the graph. Therefore, this paper proposes a novel method called Graph Kolmogorov-Arnold Convolutional Network (GKACN) where the Kolmogorov-Arnold Network (KAN) theory based convolutional operation employed in the graph for HSIC. As a new network architecture designed to replace the traditional MLP, KAN addresses the model scaling and interpretability limitations of MLP by using learnable spline-based functions in lieu of weights, resulting GKACN fully exploits the potential of the KAN and GNN structures to facilitate the mining of more representative features. Moreover, textural features from the extended morphological attribute profile extraction are used for spatial-spectral classification. Comprehensive experiments on four distinct HSI datasets have demonstrated the feasibility and superiority of the proposed method compared to other state-of-the-art models.