A hybrid runoff generation modelling framework based on spatial combination of three runoff generation schemes for semi-humid and semi-arid watersheds

It is challenging to accurately differentiate and simulate different runoff components yielded under the saturation-excess and infiltration-excess mechanisms in semi-arid and semi-humid watersheds. Fixed model structures and runoff generation modes in most of the existing conventional hydrological models limit their ability to account for respective contributions of saturation-excess and infiltration-excess flows to flood processes. To conquer this limitation, this study developed a new flexible hybrid runoff generation modeling framework, which is named as spatial combination computing models (SCCMs) for runoff generation. SCCMs can determine the dominant runoff generation mode, either saturation-excess or infiltration-excess, on the sub-watershed level and then adopt one of the Xinanjiang (XAJ), Xinanjiang-Green Ampt (XAJG) and Green-Ampt (GA) runoff generation schemes to compute runoff generation at each sub-watershed. The logical combination of the above three runoff generation schemes led to six flavors of SCCMs. Curve number and topographic index are used for dividing a watershed into infiltration-excess dominated and saturation-excess dominated sub-watersheds. The proposed models were tested in two typical semi-humid (Dongwan) and semi-arid (Zhidan) watersheds. The multi-objective generalized likelihood uncertainty estimation method (GLUE) was used to analyze the parameter uncertainty of SCCMs. The results show that all two-scheme SCCMs outperform any of the single-scheme SCCMs. SCCM-2c (a combination model of XAJ and XAJG) achieves the best performance in the semi-humid Dongwan Watershed, while SCCM-5c (a combination model of XAJG and GA) performs best in the semi-arid Zhidan Watershed. The multi-objective GLUE-based uncertainty analysis shows that parameter uncertainties of the two-scheme SCCMs (SCCM-2c and SCCM-5c) are slightly lower than those of the single-scheme SCCMs (SCCM-1n, i.e., the XAJ model, and SCCM-6n, i.e., the GA model), indicating that increased model complexity in SCCMs does not cause increased model parameter uncertainty. Our results indicate that SCCMs with more realistic representation of spatial heterogeneity can achieve higher simulation accuracies and even lower uncertainties than the conventional single-scheme models.