Hierarchical sensitivity analysis for a large-scale process-based hydrological model applied to an Amazonian watershed

Journal Article
Hydrology and Earth System Sciences, vol. 24, iss. 10, pp. 4971-4996, 2020
Haifan Liu, Heng Dai, Jie Niu, Bill X. Hu, Dongwei Gui, Han Qiu, Ming Ye, Xingyuan Chen, Chuanhao Wu, Jin Zhang, William Riley
Abstract. Sensitivity analysis methods have recently received much attention for identifying important uncertainty sources (or uncertain inputs) and improving model calibrations and predictions for hydrological models. However, it is still challenging to apply the quantitative and comprehensive global sensitivity analysis method to complex large-scale process-based hydrological models (PBHMs) because of its variant uncertainty sources and high computational cost. Therefore, a global sensitivity analysis method that is capable of simultaneously analyzing multiple uncertainty sources of PBHMs and providing quantitative sensitivity analysis results is still lacking. In an effort to develop a new tool for overcoming these weaknesses, we improved the hierarchical sensitivity analysis method by defining a new set of sensitivity indices for subdivided parameters. A new binning method and Latin hypercube sampling (LHS) were implemented for estimating these new sensitivity indices. For test and demonstration purposes, this improved global sensitivity analysis method was implemented to quantify three different uncertainty sources (parameters, models, and climate scenarios) of a three-dimensional large-scale process-based hydrologic model (Process-based Adaptive Watershed Simulator, PAWS) with an application case in an ∼ 9000 km2 Amazon catchment. The importance of different uncertainty sources was quantified by sensitivity indices for two hydrologic outputs of interest: evapotranspiration (ET) and groundwater contribution to streamflow (QG). The results show that the parameters, especially the vadose zone parameters, are the most important uncertainty contributors for both outputs. In addition, the influence of climate scenarios on ET predictions is also important. Furthermore, the thickness of the aquifers is important for QG predictions, especially in main stream areas. These sensitivity analysis results provide useful information for modelers, and our method is mathematically rigorous and can be applied to other large-scale hydrological models.