Towards global estimates of reservoir nitrogen fixation: insights from machine learning

Reservoir construction is rapidly expanding worldwide, yet their role in the global nitrogen cycle remains poorly understood. Biological nitrogen fixation (BNF) in inland fresh waters is a major pathway of nitrogen fixation but continues to be overlooked for estimates of global BNF. Here, we use a dataset of BNF in 524 reservoirs with corresponding environmental predictor variables to train and test six machine learning based models. The support vector machine model exhibited the strong predictive performance (R² = 0.71, RMSE = 0.41) and was used for further analyses. Using shapley additive explanations, we found that reservoir characteristics (volume, surface area) and climatic factors (precipitation seasonality, mean annual precipitation, and temperature seasonality) most influenced BNF rates. Notably, small reservoirs (<0.001 km³) exhibited the highest BNF rates but contributed minimally to total fixation, while a few very large reservoirs (>1 km³) dominated the total nitrogen budget due to their vast storage capacity. Applying our model to 92,527 United States (U.S.) reservoirs, we estimated an annual fixation of approximately 1.5 Tg N. Given that inland and coastal waters contribute about 15% to global BNF, reservoirs are estimated to account for 3-5% of this aquatic contribution. Within the U.S, hotspots of reservoir N-fixations were identified to be in the southeast (Florida and along the Mississippi), but they also came with the largest model uncertainties. Our findings reposition reservoirs as critical for the global nitrogen cycle, that simultaneously act as large-scale nitrogen sinks and both localized nutrient sources. We also demonstrated how machine learning can advance biogeochemical assessments at national scale, with potential for broader application.