Comparison of two model approaches in the Zambezi river basin with regard to model reliability and identifiability

Abstract. Variations of water stocks in the upper Zambezi river basin have been determined by 2 different hydrological modelling approaches. The purpose was to provide preliminary terrestrial storage estimates in the upper Zambezi, which will be compared with estimates derived from the Gravity Recovery And Climate Experiment (GRACE) in a future study. The first modelling approach is GIS-based, distributed and conceptual (STREAM). The second approach uses Lumped Elementary Watersheds identified and modelled conceptually (LEW). The STREAM model structure has been assessed using GLUE (Generalized Likelihood Uncertainty Estimation) a posteriori to determine parameter identifiability. The LEW approach could, in addition, be tested for model structure, because computational efforts of LEW are low.

Both models are threshold models, where the non-linear behaviour of the Zambezi river basin is explained by a combination of thresholds and linear reservoirs.

The models were forced by time series of gauged and interpolated rainfall. Where available, runoff station data was used to calibrate the models. Ungauged watersheds were generally given the same parameter sets as their neighbouring calibrated watersheds.

It appeared that the LEW model structure could be improved by applying GLUE iteratively. Eventually, it led to better identifiability of parameters and consequently a better model structure than the STREAM model. Hence, the final model structure obtained better represents the true hydrology.

After calibration, both models show a comparable efficiency in representing discharge. However the LEW model shows a far greater storage amplitude than the STREAM model. This emphasizes the storage uncertainty related to hydrological modelling in data-scarce environments such as the Zambezi river basin. It underlines the need and potential for independent observations of terrestrial storage to enhance our understanding and modelling capacity of the hydrological processes. GRACE could provide orthogonal information that can help to constrain and further enhance our models. In the near future, other remotely sensed data sources will be used to force modelling efforts of the Zambezi (e.g. satellite rainfall estimates) and to identify individual storage components in the GRACE observations (e.g. altimeter lake levels and microwave soil moisture). Ultimately, this will create possibilities for state updating of regional hydrological models using GRACE.