Abstract
River water temperature modeling is important for environmental monitoring and water management, especially under climate change. Recent approaches use graph-based representations to capture spatial dependencies between monitoring stations, typically based on hydrological flow. However, it remains unclear which graph structures are most suitable for representing river systems. This paper systematically evaluates the impact of various graph construction. We compare Natural Flow Graphs, Relative Neighbor Graphs, Gabriel Graphs, Delaunay Triangulation Graphs, and k-nearest Neighbor Graphs across three real-world datasets using a spatio-temporal graph neural network for air-to-water modeling. Results show that hydrological flow graphs provide a strong baseline but are not consistently optimal. Proximity-based structures often achieve lower prediction errors, with Gabriel, Delaunay, and moderately dense k-nearest Neighbor Graphs performing best. An ablation study replacing graph neural networks with multi-head self-attention achieves competitive performance but does not outperform graph-based approaches. Overall, graph construction substantially influences predictive performance.

This publication has 15 references indexed in Scilit: