Regionalization of a Distributed Hydrology Model Using Random Forest

ABSTRACT This study employs a Random Forest machine learning model in conjunction with a process‐based hydrological model for streamflow regionalisation. The models were applied across varying spatio‐temporal resolutions to test three key hypotheses: (I) the Random Forest model can better capture the nonlinear relationships between parameters and catchment descriptors compared to conventional regionalisation methods; (II) using finer temporal resolution can enhance parameter calibration, thereby improving the efficiency of the regionalisation model and (III) incorporating spatially distributed parameters can increase the model's efficiency. The results indicate that the Random Forest model outperforms conventional regionalisation methods. Furthermore, refining the temporal resolution increases model performance. For daily simulations, spatial refinement of catchment descriptors results in an approximate 10% improvement in regionalisation skill, while no discernible improvement is observed in simulations with sub‐daily time steps.