Analysis of the large-sample hydrological data set Caravan and its performance in rainfall-runoff modelling

The large-sample hydrological dataset Caravan, released in 2023, collects meteorological forcing and streamflow data from thousands of basins around the world, but has not been rigorously tested yet. Apart from the large size, the main characteristic of Caravan is the inclusion of modelled meteorological data from globally available ERA5-Land reanalysis, to ensure homogeneity among different world regions and the potential to be easily extended to additional datasets, though a cloud-based application. While these are surely great advantages, questions arise on which are the differences between the Caravan forcing data and the observation-based regional datasets, and more importantly how much they can influence the efficiency in rainfall-runoff modelling. Therefore, in the present study, the Caravan data referring to 8 national and regional data sets, including more than 3000 basins, have been first processed and analysed in terms of spatial and temporal coverage. The data are then used to calibrate the HBV-light model, evaluating the streamflow simulation performance with a two-period split-sample test. A final analysis has been devoted to the comparison of streamflow simulations in the same region (and in particular in the US) obtained when using as forcing meteorological variables either the Caravan ones or the original time-series, provided by the regional data services, in order to understand how much the ERA5-derived forcing variables influence the model simulation performance.