Hydrological modeling is used to estimate the states and fluxes in a system given inputs. This allows us to gain an understanding of the water system and make predictions such as flooding or drought. Models use parameters to describe a general process such as infiltration. These parameters can be measured, inferred or calibrated. Parameters represent local aspects such as infiltration into the soil and can differ for a catchment in rural Ohio versus mountainous Colorado. Even with the best local parameters, models will always fall short of reality since they contain simplifications. Data assimilation allows the uncertainty in the prediction of stream flow to be reduced or quantified. This thesis aims to answer the research question ’What do the adjustments in parameters and states imposed by data assimilation say about the deficiencies in the hydrological models of observed streamflow?’.

In this thesis, a data assimilation framework was developed for the eWaterCycle platform. This framework takes care of the intricacies of data assimilation for the user. Any model implemented with the Basic Model Interface can make use of this, models within eWaterCycle are focused on FAIR hydrology. The framework was implemented in such a way that any data assimilation scheme can be used, following the FAIR principles. The framework was verified to work with synthetic observations for both the HBV (Hydrologiska Byråns Vattenbalansavdelning) and the Lorenz model. The framework also makes applying data assimilation within eWaterCycle easier and scalable.

To answer the research question, the five-year hydrological response of 671 catchments in the United States of America was modeled. These catchments are from the CAMELS dataset, which is a collection of hydrological data with characteristics for catchments across the United States of America. The data assimilation scheme particle filtering was applied to the conceptual hydrological model HBV using eWaterCycle. The parameters and states of each experiment were stored and analysed. The data assimilation scheme requires hyperparameters. These hyperparameters were optimised using the first 26 catchments in the dataset. This single combination of hyperparameters was used across the 671 catchments.
Taking the best prediction of stream flow at every time step, the data assimilation experiment outperformed the calibrated model in 461 of these catchments. Part of this is due to the bias introduced through the use of one set of hyperparameters. Analysis shows that catchments of the same characteristics perform similarly. Catchments with a low mean stream flow, have a high spread of predicted observations causing very good predictions when selecting the best. Due to this bias, no real correlations can be drawn about the relation of background characteristics and model deficiency. Analysing three catchments on a catchment scale showed that step changes in parameters often occur around flood peaks. The data assimilation scheme adjusts the working of the model to better capture the observed streamflow.

To further incorporate the framework in eWaterCycle more testing should be done on 2D models. Deficiencies in the HBV model can be further analysed by optimising hyperparameters per catchment type or flow regime.

Rain gauges are a powerful tool to measure rain entering a watershed. When water flow through a watershed is modeled, these rainfall measurements are used as inputs. Hydrological models have become increasingly complex as they more accurately represent the physical processes occurring. This is mostly done by increasing the spatial and temporal resolution of the model. As this resolution is increased, the inputs also need to increase. This thesis looks into if rain gauges are in the right place when used as inputs for hydrological models. This has been done by analysing four factors which literature showed to have affect rain gauges. The four considered factors are: the distribution of rain gauges, the steepness of the slope they are on, the location on that slope and their location within a watershed. For each of these factors algorithms have been developed in Python which compute relevant information on a given station. These algorithms have been applied to 368 gauges across the United Kingdom (UK), available from an open data source. The rain gauges are well distributed across different altitudes matching the distribution of heights across the UK. Above 400 m there are no gauges and this area is therefore underrepresented. The spacing of stations is good, a few close together and some isolated gauges on islands. The steepness of slopes varies strongly, when a steepness of 25% is used as a threshold only around 3% are on too steep of a slope. A fair amount of gauges are on ridges. Especially those near the coast have steep seaward slopes and thus will suffer from underestimating the actual rainfall. Within watersheds gauges are often near rivers causing other areas of the watershed to be underrepresented, especially areas of higher elevation. In future research it is recommended to use more gauges in the data set. Secondly focusing on a baseline comparison can help identify which stations are placed incorrectly. Lastly it is recommended to vary the resolution of elevation data and the spatial area considered, focusing on watersheds.