Modelling Brazilian hydrology using various input datasets and model structures

Large-sample hydrological datasets and increasing computational power allow us to conduct modelling studies that were previously impossible. For example, it is now possible to test how different model structures affect the simulated streamflow dynamics and model performance for a variety of catchments instead of only a handful (but well-known) catchments. While this is excellent progress, the modeller generally does not understand the physical processes nor the reliability of the data for the hundreds of catchments as well as for the limited number of catchments.

The two Brazilian large-sample hydrology datasets CAMELS-BR and CABra were created for the same purpose but differ in content as they are based on different types of meteorological data. CAMELS-BR is mainly based on large-scale data from satellite, reanalysis, and gauge data, while CABra is based on the interpolation of station data. Especially the potential evapotranspiration values differ strongly for the two datasets, with the annual sums in CAMELS-BR being only 50 to 70 % of those in CABra. This situation enables us to test if different model structures enhance process representation or mainly compensate for flawed input data.

We tested the two datasets with three versions of the HBV model. Aside from the standard version simulating soil moisture and evapotranspiration, percolation, and streamflow from groundwater, we used a model version that can accommodate inter-catchment groundwater flow (i.e., inflow or outflow of groundwater), as well as a simpler version of the model in which a constant part of the precipitation is assumed to become groundwater and the remaining part is not explicitly included (i.e., soil moisture and evapotranspiration are not explicitly simulated). Although evapotranspiration is represented in a very simplified manner, this has the advantage that no (uncertain) potential evapotranspiration data are required. Interception losses can be represented better as they are not part of the evapotranspiration from the soil routine.

Our results show that large-sample datasets are very useful for testing different model structures and thus representation of hydrological processes. Regardless of the dataset used, the model version has a large effect on the streamflow simulations. The best results were usually achieved with the simplified soil routine, even though it has fewer parameters that need to be calibrated. Allowing for intercatchment groundwater flow improved the performance compared to the standard version of the model in many cases as well.