Selecting model formulations and parameterizations: Reducing the need for calibration using open data and landscape characteristics

Abstract

The need for calibration of conceptual hydrological models on river discharge is still large, and the scope of this research is to reduce this need based on new parameter estimation techniques, additional information sources and hydrological understanding. In a first step, a regularized model and an adjusted regularized model with sub-grid variability based on the landscape, both constrained and unconstrained, were calibrated for seven catchments. Four catchments were also simultaneously calibrated and their feasible parameters transferred to the remaining three receiver catchments. Small improvements were observed by introducing sub-grid variability, whereas the semi-quantitative constraints led to moderate improvements compared to the unconstrained model. Especially low flow statistics improved and suitable prior constraints can aid model transferability. Subsequently, it was assessed how well the key parameter of root-zone storage capacity could be estimated in three deforested catchments. A recently introduced method based on rainfall and an estimate of transpiration was used to reproduce the temporal evolution of root-zone storage capacities. These values were compared to the values from four hydrological models calibrated for consecutive 2-year windows. Water-balance derived root-zone storage capacities showed a similar signal compared to the calibrated values of the models with a sharp decline in root-zone storage capacity after deforestation, followed by a gradual recovery of 5 to 13 years. The added value of several combinations of remotely sensed products for parameter estimation was assessed for five different hydrological models in 27 catchments across Europe. A parameter selection process was applied for 1023 possible combinations of ten different data sources, ranging from using 1 to all 10 of these products. High probabilities of improvement, with regard to commonly applied model performance metrics, were obtained when combinations included AMSR-E and ASCAT soil moisture, and GRACE total water storage anomalies. The evaporation products of LSA-SAF and MODIS were less effective for deriving meaningful posterior parameter distributions. In a last step, a large-scale hydrological model with landscape-derived sub-grid variability was run with 50 random parameterizations for the European continent, with and without semi-quantitative prior parameter constraints. A variable pattern in improvements/deteriorations was observed when evaluated for 397 gauging stations, which shows that the prior parameter constraints were not sufficient to limit the search space effectively. Concluding, this thesis presents several ways to better optimize different hydrological models without using observed river discharge for varying spatial scales and changing circumstances. In this way, this thesis serves as a stepping stone towards fully predicting in ungauged catchments.