Hybrid AI for estimating water levels at ungauged river locations via graph neural networks and terrain-aware edges

Abstract

Monitoring river water levels is critical for flood risk management, water-resource planning, and early warning systems. However, deploying dense gauge networks across extensive river systems is often infeasible due to logistical and financial constraints, and existing stations may fail or provide intermittent data. In this work, we propose HIGNN (Hydrological Interpolation based on Graph Neural Network), a graph-based framework for estimating water-level changes at virtual sensor locations (i.e., ungauged sites or locations with missing observations) by leveraging information from neighboring telemetry stations and terrain characteristics. In HIGNN, nodes represent observation sites, waterway intersections, or virtual stations, while edges represent hydrological connectivity (e.g., upstream–downstream relations) and are characterized by topographic attributes such as elevation profiles, slope statistics, and flow direction. The model employs message passing to propagate water-level change signals through the river network, modulated by physically meaningful edge attributes. Across all water-level change brackets, HIGNN achieves the lowest mean RMSE, outperforming interpolation- and regression-based baselines. These results demonstrate that HIGNN can effectively estimate water-level changes at ungauged or temporarily unmonitored locations.