Improving the performance of distributed conceptual hydrological models using the spatio-temporal patterns of RS observations

Improving the performance of distributed conceptual hydrological models using the spatio-temporal patterns of RS observations

More Info
expand_more

Abstract

Hydrological models are used for all kinds of water management applications. Detailed hydrological simulations are needed to solve the hydrological problems of the 21st century, especially in developing countries. However, sufficient hydrological and meteorological data is often not available. The use of remote sensing (RS) datasets may offer a solution to this problem. RS datasets can perfectly be applied in distributed conceptual hydrological models. In this study, several RS datasets are applied in the calibration of a distributed conceptual hydrological model, and the influence of this approach on the overall model performance is assessed. The RS data applied in this study include terrestrial water storage anomaly (TWSA) data, normalized difference vegetation index (NDVI) data, and soil moisture (SM) data. Also the input and forcing data for the hydrological model consists of datasets based on satellite observations. This data is used in a wflow hbv model, which is applied to the Volta basin in Western Africa as a case-study. In this study, not only the effect of including RS data in the calibration of a distributed hydrological model on streamflow is assessed, but also the effect on a set of internal components of the system, directly related to the datasets used for calibration. These internal stocks and fluxes are the TWSA, the actual evapotranspiration (AET) and the amount of soil moisture in the unsaturated zone. Together with streamflow, the assessment of these stocks and fluxes make up the overall model performance of the system. The effect on the overall model performance is examined using different scenarios, in which different combinations of datasets are used for calibration. Not the absolute values, but the spatiotemporal patterns of the remote sensing datasets are used for model assessment. This is done using the spatial pattern efficiency metric (ESP ). Model optimization was done using the Dynamically Dimensioned Search (DDS) algorithm. The results show that the hydrological model developed for this case-study is already able to simulate streamflow and the temporal patterns of the RS datasets quite well, when it is calibrated on streamflow only. However, the spatial pattern representation of the RS datasets was found to be inadequate and the differences in streamflow simulation performance for the different subcatchments is large. When SM or TWSA data was added to the calibration procedure, the temporal and spatial pattern representations only changed minimally, which is attributed to limited model complexity and flexibility. However, generally a trade-off effect was observed in which the spatial and temporal pattern representation improved, but the streamflow performance decreased. This effect was stronger for the addition of the SM dataset to the calibration than for the addition of TWSA dataset. Although there is definitely a strong connection between NDVI and AET, the physical relation between the two variables was found to be too weak to be used for hydrological model calibration, even when only the spatial and temporal pattern information was used. The overall model performance did improve most in the calibration catchments in the scenario in which Q, SM and TWSA data were combined in the calibration procedure, but the differences with the baseline scenario were only small. For the streamflow performance however, the differences between the scenarios are quite significant. It was shown that calibrating a hydrological model on the spatial and temporal patterns of RS data only (non-Q calibration) can accurately represent the temporal pattern of streamflow observations, but not the magnitude of the flow values. It is recommended to repeat this study using a more complex and more flexible model setup, which allows the model to use the freedom it is given to better represent the spatial patterns observed with RS.