Advances in graph neural network-based hyperspectral remote sensing image classification

The rapid development of remote sensing technology has brought a variety of remote sensing data. Hyperspectral images, with the highest spectral resolution among these data, are always a crucial source for various Earth observation applications. In the field of computer vision, pattern recognition algorithms represented by deep learning also constantly developing and breaking through the limitations, providing more effective technologies for hyperspectral remote sensing applications. In recent years, Graph Neural Networks (GNNs) have been widely utilized in hyperspectral remote sensing image interpretation tasks, which can leverage the underlying relationships between samples to extract both local and global contextual information, producing high-precision classification results even with a limited number of labeled samples. This paper summarizes the most commonly used GNN frameworks from existing studies, analyzing the characteristics of these methods by decomposing their structures and categorizing them. We first extract the commonalities of existing GNN architectures and propose a basic module of GNN, which consists of an information aggregation function and a feature updating function. Building upon this module, we reinterpret various popular GNN architectures, including spectral-based Graph Convolutional Networks (GCNs), spatial-based GCNs, and Graph Autoencoders (GAEs). In the context of GAEs, current approaches are analyzed from three perspectives: loss functions, decoders, and graph reconstruction methods. These methods formulate loss functions to incorporate various graph-based constraints, thereby embodying the implicit assumptions and specific characteristics inherent to each method. Then, the analyses of GNN methods in the remote sensing field are conducted from three perspectives: graph connections, graph nodes, and network models. The existing research outcomes are classified based on the spatial range of connections, the information hierarchy of nodes, and the uncertainty of models. These GNN-based algorithms extract either local information or non-local information (e.g., global or local-global interactions) by using graph connections across different spatial ranges. The concept of non-local modeling has been extensively explored in GNNs over the past four years. Among GNNs with different node information hierarchies, approaches of using superpixels as graph node representations are the most prevalent. This is because superpixels can serve as a generalized form of node representation for other hierarchies, and their construction is relatively straightforward. Additionally, this paper introduces the application of GNNs in hyperspectral remote sensing image classification under varying modal and label quantities. For single-modal applications, we summarize the characteristics of several representative algorithms and provide their corresponding code implementations. For a limited number of multi-modal applications, we categorize and introduce methods based on the role of GNNs in multi-modal feature fusion. We conduct detailed analyses of the performance of GNN-based classification models in relevant literature, evaluating the applicability of these methods by considering the number of labeled samples and their corresponding classification accuracies. Furthermore, we elaborate on the theoretical foundations and integrated techniques of these models in the fields of supervised, semi-supervised, and unsupervised classification. Finally, the paper summarizes and looks forward to the frontier technologies of GNNs from three aspects: deep graph networks, GNN integrated with other deep learning techniques, and GNN-based foundation models, providing directions and insights for future research in the remote sensing field. © 2025 Science Press. All rights reserved.