Deriving hydrological inferences from a machine learning model to understand the physical drivers of flow duration curves

• An explainable ML approach to identify key drivers of FDCs is presented. • The RF model adequately predicts FDCs in ungaged watersheds. • SHAP revealed global, regional and local controls on FDCs • Aridity Index and annual precipitation drive the scale of FDCs. • Baseflow Index, Seasonality Index, and geologic attributes control FDC shape. Flow Duration Curves (FDCs) provide a statistical relationship between the magnitude and frequency of streamflow in a watershed. The shape and scale of an FDC are dependent upon the dynamic relationship between streamflow regimes and the climatic, geological, topographical, and other environmental or anthropogenic attributes of the watershed. While Machine Learning (ML) models can provide enhanced predictive performance in ungaged watersheds over conventional approaches, validating the models’ comprehension of the underlying hydrologic processes is imperative for any stakeholder involvement. In this study, we employed Random Forest (RF) regression on individual Exceedance Percentiles (EPs) and slope of the FDCs for a large sample of watersheds in the contiguous United States. Then, we explored the interactions between watershed attributes and FDCs using SHapley Additive exPlanations (SHAP) to divulge the local (watershed scale), regional, and global (model scale) influence of attributes and their dependence structures. Results indicate that climate attributes (precipitation and aridity index) were the preeminent drivers in predicting FDCs across all quantiles, primarily affecting the scale of FDCs, followed by the baseflow index and geologic attributes that highly influence the low flow regime and control the shape of FDCs. The dominant controls of other watershed attributes, including precipitation seasonality, % snow, and elevation, vary between EPs and regions that were readily discernible in their SHAP values.