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

Area of Science:

• Hydrology
• Water Quality Modeling
• Machine Learning

Background:

• Accurate prediction of streamflow and nitrogen export is crucial for agricultural watershed management.
• Existing temporal deep learning models (e.g., LSTM) struggle with spatial generalization due to limited representation of spatial dependencies and flow paths, particularly in data-scarce regions.

Purpose of the Study:

• To introduce HydroGraphNet, a novel knowledge-guided graph machine learning framework for spatially distributed hydrological and biogeochemical predictions.
• To enhance the spatial generalization capabilities of deep learning models in data-scarce agricultural watersheds.

Main Methods:

• Developed HydroGraphNet, integrating process-based knowledge and explicit spatial learning into temporal modeling using directed graph topology and mass balance constraints.
• Pretrained HydroGraphNet on synthetic data from SWAT+ (Soil and Water Assessment Tool Plus) to improve generalization in sparsely monitored areas.
• Evaluated HydroGraphNet in the Upper Sangamon River Basin, comparing it against lumped and distributed LSTM baselines.

Main Results:

• HydroGraphNet significantly improved temporal and spatial extrapolation performance compared to the lumped LSTM baseline when benchmarked on SWAT+ simulations.
• After fine-tuning with USGS data, HydroGraphNet achieved superior mean test NSE (KGE) scores for discharge and NO₃-N load compared to baselines.
• Attribution analysis confirmed the importance of upstream inflow representation and graph-based spatial learning for capturing cross-subwatershed dependencies.

Conclusions:

• HydroGraphNet offers a generalizable framework for spatially distributed prediction, advancing the integration of physical knowledge and spatial learning in hydrological modeling.
• The model demonstrates robustness and process fidelity, capable of reproducing seasonal hydrological and biogeochemical patterns.
• This framework supports targeted water quality management in data-scarce watersheds.