Global runoff partitioning based on Budyko and machine learning

Partitioning runoff accurately into baseflow and quickflow is crucial for our understanding of the water cycle and for the management of droughts and floods. However, global datasets of long-term mean runoff partitioning are rare, even more so datasets relying on physically-based methods. Here, we present a new global 0.25° dataset of runoff, baseflow and quickflow using a hybrid approach of Budyko-based methods and machine learning (ML). The parameters in the Budyko curve and Budyko-based baseflow curve (BFC curve) are estimated with ML (boosted regression trees, BRT) as a function of catchment characteristics. The BRT models are trained and tested in 1226 catchments worldwide, and then applied globally at grid scale. The catchment-trained models show good performance during the testing phase with R² equal to 0.96 and 0.87 for runoff and baseflow, respectively. The dataset developed in this study shows that 30.3±26.5% (mean ± standard deviation) of the precipitation is partitioned into runoff of which 20.6±22.1% is baseflow and 9.7±10.3% is quickflow. The global long-term mean baseflow in this study (151±181 mm yr^–1) is lower than that from the Global Streamflow Characteristics Dataset (GSCD, 241±321 mm yr^–1) and higher than that from ERA5-Land (79±145 mm yr^–1). This study provides a unique, physically and observationally constrained global dataset of the long-term runoff partitioning. The large differences among different datasets suggest that global runoff partitioning is highly uncertain and requires further investigation.