Evaluating the performance of random forest for large-scale flood discharge simulation

The machine learning algorithm ‘random forest’ has been applied in many areas of water resources research including discharge simulation. Due to low setup and operation cost, random forest could represent an alternative approach to physical and conceptual hydrological models for large-scale hazard assessment in multiple catchments. Yet, the applicability of random forest to flood discharge simulation requires further exploration, especially with respect to heterogeneous catchments and daily temporal resolution. In this study, we simulate flood event and peak discharge on a daily time scale for 95 study basins in Canada and the USA. We comparatively evaluate the predictive performance of random forest against the conceptual hydrological modeling package ‘hydromad’ and assess the influence of catchment characteristics on model performance. Our analysis showed that random forest is competitive to hydromad in the simulation of low and medium flood magnitudes. However, both models exhibit inaccuracies for higher flood events. Relating catchment characteristics to model skill, we found that primarily climatic conditions and elevation affect the flood simulation capability. We conclude that random forest provides a low-cost and, yet, competitive alternative to conventional rainfall-runoff models in large-scale flood discharge simulation. Nevertheless, without further model advancements, the presented models only provide robust discharge predictions for small and medium magnitude floods in low altitude catchments with warm temperate climate.