Edge-of-field runoff prediction by a hybrid modeling approach using causal inference

Abstract

Unforeseen runoff events cause nutrient losses that affect crop production, revenue, and contribute to deteriorated water quality, leading to harmful algal blooms and hypoxia in receiving water bodies in the Great Lakes region. To mitigate the negative impacts caused by runoff events, we developed a hybrid modeling approach by combining physics-based and statistical models to predict the occurrence and level of severity of daily runoff events, supporting agricultural producers to avoid nutrient application before significant runoff events. We chose to use the National Oceanic and Atmospheric Administration’s National Water Model (NWM) as the physical model given its flexible architecture design, technical robustness, model resolution, data availability, and wide application scale. For the statistical model, we developed a data-driven tool built from Directed Information and eXtreme Gradient Boosting (XGBoost) to estimate the occurrence and the level of severity of daily edge-of-eld (EOF) runoff events. This data-driven tool ingests a large variety of variables from NWM operational runs and translates them into the EOF runoff predictions on a daily scale in the Great Lakes region. Without calibrating the large-scale NWM for the local runoff prediction, the results show large improvements in the prediction of the occurrence and level of severity of daily EOF runoff using the hybrid physical-statistical modeling approach. Ultimately, the hybrid approach, when integrated into runoff risk decision support tools, is expected to provide dual benefits to agricultural producers and water quality, retaining more nutrients on their fields and lowering nutrient loads to water bodies during runoff events.