Towards operational hydrological model calibration using streamflow and soil moisture measurements

Abstract

Timely and reliable flood forecasting is critical for flood warning delivery and emergency response. As core components of an operational forecasting system, hydrological models are typically calibrated using streamflow measurements to minimize parameter uncertainties. The rapid development of earth observation techniques provides opportunities to obtain soil moisture information. As catchment discharge is strongly related to soil moisture, there is a possibility to improve streamflow forecasts by using soil moisture measurements for model calibration. The use of soil moisture observations has attracted increasing attention, however, there have been a limited number of studies. This study aims to assess the impact of integrating soil moisture measurements for model calibration on the forecast skills of hydrological models. Experiments were implemented in the Adelong Creek (157 km 2 ) and the Upper Kyeamba Creek (190 km 2 ) catchments of the Murrumbidge Basin using a lumped rainfall-runoff model named GRKAL (modele du Genie Rural Kal). Two calibration scenarios are performed: 1) a traditional streamflow-only-calibration; 2) a joint-calibration using both streamflow and in-situ soil moisture measurements. Outcomes are evaluated in a hind-casting mode for both a calibration and independent validation period. Results show that, for the Adelong catchment, the joint-calibration led to a slightly worse match between the simulated and observed streamflow with a Nash-Sutcliffe (NS) value of 0.8173, as compared to the streamflow-calibration scheme (achieved a NS value of 0.8443) alone in calibration period. During the validation period, the joint-calibration achieved a NS value of 0.7952, performing better than the streamflow- calibration scheme which gives a NS value of 0.7586. This result indicates that although introducing the soil moisture measurements to the objective function lead to a sub-optimal match of simulated streamflow to the observed data during the calibration period, there exists the possibility that joint calibration potentially optimized the model parameters to be more realistic, resulting a more precise prediction in the validation period. However, for the Kyeamba catchment, the results tend to be worse: NS values derived from joint- calibration was less than that obtained by streamflow-calibration in both calibration and validation period. This may relate to the unphysically based model structure, equal-weighted objective function, and various sources of uncertainties. In terms of the soil moisture prediction, it is consistent in both catchments that the joint-calibration illustrates a marginally better match to the observed value than the streamflow calibration during most of the study period. It is concluded that while the joint-calibration will typically lead to poorer streamflow forecast results during the calibration period, it could lead to a more robust result in the validation/forecasting period. As this was not consistently the case, more objective functions (such as unequal-weighted NS and a combination of several frequently-used objective functions) need to be investigated to identify the best calibration strategy for using soil moisture information.