Floods, droughts, and heatwaves are increasing globally. This is often attributed to CO₂-driven climate change. However, at the global scale, CO₂-driven climate change neither reduces precipitation nor adequately explains droughts. Land-use change, particularly soil sealing, compaction, and drainage, are likely more significant for water losses by runoff leading to flooding and water scarcity and are therefore an important part the solution to mitigate floods, droughts, and heatwaves.

The Unsolved Problems in Hydrology (UPH) initiative has emphasized the need to establish networks of multi-decadal hydrological observatories to tackle catchment-scale challenges on a global scale. This opinion paper provocatively discusses two end members of possible future hydrological observatory (HO) networks for a given hypothesized community budget: a comprehensive set of moderately instrumented observatories or, alternatively, a small number of highly instrumented super-sites.

We compare the catchment forcing data provided in large-sample datasets, namely the Caravan dataset and three of the original CAMELS datasets (US, BR, GB). We show that the differences affect hydrological model performance and that the data quality in the Caravan dataset is lower than the one in the CAMELS datasets, both for precipitation and potential evapotranspiration. We want to raise awareness of the lower data quality in Caravan and we suggest possible improvements for the Caravan dataset.

Recently, a special type of neural network architecture became increasingly popular in hydrology literature. However, in most applications, this model was applied as a one-to-one replacement for hydrology models without adapting or rethinking the experimental setup. In this opinion paper, we show how this is almost always a bad decision and how using these kinds of models require the use of large-sample hydrology datasets.