Knowledge-guided graph machine learning for spatially distributed prediction of daily discharge and nitrogen export dynamics

Spatially distributed prediction of streamflow and nitrogen (N) export dynamics is essential for precision management of agricultural watersheds. While temporal deep learning models have shown strong basin-scale performance, their ability to generalize spatially is limited, particularly under data-scarce conditions. To address this gap, a team of researchers led by the Center for Advanced Bioenergy and Bioproducts Innovation (CABBI) propose HydroGraphNet, a knowledge-guided graph machine learning framework integrating process-based knowledge and explicit spatial learning into temporal modeling.

The HydroGraphNet framework incorporates directed graph topology to encode watershed connectivity and upstream inflows, using mass balance constraints to improve physical consistency. It was pretrained on synthetic data to enhance generalization in sparsely monitored regions. HydroGraphNet was evaluated in the upper Sangamon River Basin against two baselines.

After fine-tuning the model with USGS monitoring data, the model substantially outperformed baseline for both discharge and NO₃–N load. Attribution analysis further highlighted the importance of upstream inflow representation and graph-based spatial learning in capturing cross-subwatershed dependencies. The model reproduced seasonal hydrological and biogeochemical patterns consistent with known processes, demonstrating its robustness and process fidelity for spatially distributed predictions.

HydroGraphNet offers a generalizable framework for distributed modeling to support spatially targeted water quality management in data-scarce watersheds.

This work was funded in part by CABBI, a U.S. Department of Energy-funded Bioenergy Research Center, with a grant from the DOE Office of Science, Biological and Environmental Research Program.

Images available upon request.