Graph neural networks (GNNs) are neural models designed for learning on graph‑structured data. Applying them to electronic navigational charts (ENCs), which are geospatial vector datasets, is a natural step, yet it remains unclear which message‑passing scheme, architecture, and graph representation best serve the classification of change criticality. In our approach, spatial objects in an ENC become graph nodes, while their spatial proximity and semantic relationships form edges, yielding a unified graph. The old ENC and the new ENC are each encoded as a graph, and the problem is cast as a graph‑pair classification task: does the difference between the two graphs represent a critical safety risk? Building on this representation, we evaluate several mainstream GNN architectures (e.g., GCN, GraphSAGE, GAT) on a dataset of changes reviewed by maritime experts. The experiments demonstrate that graph‑based representations markedly improve the accuracy of distinguishing critical from non‑critical updates, offering a scalable route toward automating ENC maintenance workflows.
Review: The work showcases the effectiveness of GNNs for spatial change detection and provides a solid technical foundation for intelligent management of maritime chart updates.