A few hundred catchments later – lessons learned from modeling large catchment samples

Traditionally, hydrological models are applied to one or a few catchments because preparation of the input and calibration data for a more extensive set of catchments is challenging. The availability of data sets with hydrometeorological time series for large numbers of catchments has been a game changer in hydrological catchment modeling in recent years. One example are the CAMELS data sets with the basic data to run hydrological models for several hundreds of catchments in various countries. In several recent studies, we have used these data sets for bucket-type modeling of a large number of catchments in different regions. In this presentation, I will discuss some of our main findings:

• Variability of results: Simulation results vary considerably between catchments, making it pertinent to apply a model to a large number of catchments for robust results.
• Uncalibrated model performance: Simple bucket-type models can provide surprisingly good results for some catchments even when not calibrated. This needs to be considered when we assess model performances.
• Prediction in ungauged catchments: It can be challenging to improve simulations for ungauged catchments by regionalization as it is not obvious how to choose the most suitable donor catchments. Thanks to data sets with a vast number of potential donor catchments, we found that almost perfect donor catchments seem to exist in most cases. However, the challenge remains to identify them.
• Model structure: For some catchments, a simplified soil routine with only one free parameter (instead of three) outperformed the standard model version.
• Value of data: Large samples of catchments allow us to evaluate the value of different data types: a limited number of streamflow gaugings and other data types, such as stream level, stream width or water level class data, can be informative for streamflow simulations.