Evaluation of random forests for short-term daily streamflow forecasting in rainfall- and snowmelt-driven watersheds

Abstract. In the past decades, data-driven machine-learning (ML) models have emerged as promising tools for short-term streamflow forecasting. Among other qualities, popularity of ML such applications is due to their relative ease in implementation, less strict distributional assumption, and competitive computational predictive performance. Despite encouraging results, most forecasting been limited watersheds which rainfall major source runoff. this study, we evaluate potential random forests (RFs), a popular method, make forecasts at 1 d lead time 86 Pacific Northwest. These cover diverse climatic conditions physiographic settings exhibit varied contributions snowmelt streamflow. Watersheds are classified into three hydrologic regimes based on timing center-of-annual flow volume: rainfall-dominated, transient, snowmelt-dominated. RF performance benchmarked against naïve multiple linear regression (MLR) evaluated using four criteria: coefficient determination, root mean squared error, absolute Kling–Gupta efficiency (KGE). Model evaluation scores suggest that performs better snowmelt-driven compared rainfall-driven watersheds. The largest improvements benchmark found among deteriorates with increases catchment slope soil sandiness. We note disagreement between two measures variable importance recommend jointly considering these physical processes under study. results presented provide new insights effective application RF-based