Inferring hydrological process understanding using models and large-sample data sets
Catchments are complex systems, which have evolved under the influence of environmental processes over long periods of time. Due to this inherent complexity, it is often difficult to understand what forces the hydrological behavior of a given catchment. The two most common approaches to better understand catchments are creating hydrological models ... to test which hypothesis of catchment functioning works best or to look at the catchment characteristics and try to infer the most important forcing directly from this. This dissertation uses both approaches to reach a more holistic understanding of catchment functioning. The starting point of this dissertation was an earlier publication of mine, which used an innovative way for model development, the so called “incremental model breakdown”. This new approach starts with a complex model, incrementally deactivates processes and checks which process deactivation causes the model to fail. All processes that lead to a failure when deactivated show that they are important for the model. This enables a more thorough exploration of the space of possible model structures than traditional approaches, which start from predefined structures. However, during the development of this approach it became apparent that model parameters are able to compensate extensively for the omission of processes. Therefore, larger and well-understood data sets are needed to form hypotheses of catchment functioning that could be tested by incremental model breakdown. Based on this prior knowledge and to lay the foundation for future research, this dissertation builds on the incremental model breakdown approach and examines two large sample data sets with different methods. As the main problem of the new model building approach was the way it handles the model parameters, the first part of this dissertation focusses on the intricate interaction between model complexity and parameter uncertainty. This is done by exploring the trade-offs between tightly constrained parameters and the ability of the hydrological models to predict hydrological signatures that capture the behavior of a river. The results show that there is a clear trade-off along the axis of complexity for those models. The simpler a model is, the better its ability to constrain parameters, but the worse are the results of an independent validation of its realism using hydrological signatures. Those results highlighted again that hydrological models can only be as good as the hypothesis forming their basis and those hypotheses can only be found and improved by looking at real catchments’ data. These datasets need to contain the hydrological behavior and characteristics of catchments to facilitate deriving hydrological process understanding – and develop appropriate models that reflect this catchment’s understanding. Therefore, in the second part of my thesis, is about the exploration of two large hydrological datasets. The first step was an analysis of the CAMELS dataset, a large-scale open access dataset that contains catchments from all over the continental United States. This allows to determine the most important factor for the overall hydrological behavior, namely the climate, and more specifically, the aridity and the frequency of large precipitation events. However, the results also show that this climatic forcing can be found more directly in some catchments than in others. This was likely a problem of scale, given the continental domain of this study. To better understand why this is the case, we established a second dataset which only contained catchments from Hesse, Germany, for the third part of my work. This allowed looking at how catchments with different characteristics behave under the same climatic forcing. The focus here was the complexity of the storage-discharge relationship. The results showed that the hydrological signal of the climatic forcing is mainly influenced by the catchment’s permeability, conductivity, geology, soil and, to a lesser extent, its topography. It also showed that the complexity of the hydrological response differs strongly between catchments. While some catchments show a storage-discharge relationship that is almost exactly an exponential function, others show a more erratic behavior. The properties of the simple catchments all facilitate a higher interconnectedness of the storage system of the catchment; this indicates that the complexity of a catchment’s behavior is strongly linked to its overall connectivity of water pathways. To finally use this improved hydrological understanding and connect it to model structures, preliminary tests link the catchment complexity with modelling ease. For this, I used the simple HYMOD model and evaluated its performance for catchments of differing complexity. Those results showed that the simplicity in behavior is connected to the model performance. The simpler the storage-discharge relationship, the simpler the model for the catchment can be. The findings of this dissertation highlight that even though it is possible to change model structures and calibrate parameters to get results with high values for the objective function, those models can still have difficulties in independent verification. This indicates that the models are often “right for the wrong reasons”. Only if we thoroughly understand datasets that capture a wide variety of hydrological behavior and catchments characteristics will we be able to construct more realistic hydrological models that are “right for the right reasons”.