Nonlinear control of climate, hydrology, and topography on streamflow response through the use of interpretable machine learning across the contiguous United States

Abstract

Abstract Runoff has been greatly affected by climate change and human activities. Studying nonlinear controls on runoff response is of great significance for water resource management decision-making and ecological protection. However, there is limited understanding of what physical mechanisms dominate the runoff response and of their predictability over space. This study analyzed the spatial patterns of runoff response including runoff changes and its sensitivity to climate–landscape variations in 1,003 catchments of the contiguous United States (CONUS). Then, an interpretable machine learning method was used to investigate the nonlinear relationship between watershed attributes and runoff response, which enables the importance of influencing factors. Finally, the random forest model was employed to predict runoff response according to the predictors of catchment attributes. The results show that alteration of runoff is up to 56%/10 years due to climate change and human activities. Catchment attributes substantially altered runoff over CONUS (−60% to 56%/10 years). Climate, topography, and hydrology are the top three key factors which nonlinearly control runoff response patterns which cannot be captured by the linear correlation method. The random forest can predict runoff response well with the highest R2 of 0.96 over CONUS.