Drivers of nitrate removal rates in stream restoration: a machine learning assessment for watershed-scale analysis

Eutrophication of inland and coastal waters from excess nutrients impacts human and ecosystem health. River restoration is used to reverse such impacts but the effects of multiple restoration projects at large temporal and spatial scales are poorly understood. A key need is understanding controls on restoration effects, particularly controls relevant at larger spatial and temporal scales such as Strahler stream order, hydrologic flow condition, and season of the year. We conducted a literature review to gather nitrate removal rates from field studies of stream restoration projects that included both ‘actual’ (measured in-situ under field conditions) and ‘potential’ (maximum values measured ex-situ in labs) rates. We then used a random forest (RF) machine learning approach to determine which controls are most important, and how predicted nitrate removal varies with key controls. Literature review results indicate that field data were relatively sparse, with studies heavily concentrated in smaller streams, highlighting the need for systematic data from larger watersheds. RF results for actual rates indicate that hydrologic condition was the most important control, confirming the importance of lateral connectivity in river systems to excess nitrate processing. Mean predicted nitrate removal rates were higher for baseflow than stormflow conditions, consistent with differences in residence time. Steam order was of moderate importance, with predicted nitrate removal rates higher in smaller streams. Overall, nitrate removal rates were variable in space and time, underscoring the importance of analyzing the effects of stream restoration on watershed nitrogen budgets in spatially and temporally explicit ways at watershed scales.