Comparison and integration of hydrological models and machine learning models in global monthly streamflow simulation

With the advancement of machine learning technology, machine learning models (MLMs) are progressively emerging as a pivotal branch in streamflow simulation. However, compared to traditional hydrological models (HMs), a comprehensive understanding of their strengths and applicability across diverse climatic regions worldwide remains elusive. This study compares the performance of four widely used lumped HMs (GR2M, XAJ, SAC, and Alpine) with four prevalent MLMs (RF, GBDT, DNN, and CNN) across 16,218 catchments worldwide. Results show that MLMs can't always surpass HMs. Although the percentage of qualified-level accuracy (Kling–Gupta efficiency, KGE ≥ 0.2) is higher in the MLMs group, the HMs have a higher percentage of good-level accuracy (KGE > 0.6). HMs outperform MLMs in Southeastern North America, Western Europe, and most regions of the Southern Hemisphere. To combine the merits of HMs and MLMs, the performance of seven multi-model weighting ensemble methods (MWEs) is evaluated. The optimal MWE is employed to unify the results of four HMs and four MLMs, further improving the simulation accuracy. The MWEs improve the simulation accuracy effectively and the Inverse Rank Prediction Combination (InvW) is the best-performing MWE, which elevates the percentage of qualified accuracy by 13 % and 6 % for the best-performing HM and MLM, respectively. Despite the improvements with InvW, correlation analyses of KGE with catchment humidity index (HI), mean temperature (Tair), and leaf area index (LAI) reveal that simulating streamflow in dry, cold, and lower vegetated areas remains challenging (R_HI = 0.36, R_Tair = 0.1, and R_LAI = 0.19). Additionally, the precision of streamflow simulations diminishes in areas heavily impacted by human activities, primarily due to hydraulic engineering and water withdrawals. Our study systematically evaluates the performance of different HMs and MLMs in different regions, proposes a framework to combine the merits of HMs and MLMs, and sheds light on the constraining factors of accurate streamflow simulation.