From Model Performance to Model Behavior through Fingerprinting of Hydrologic Models (Invited Presentation)

Hydrology as a field of study and application is characterized by the enormous proliferation in the number of models that are in use. This reflects the difficulty in describing hydrologic processes over a range of temporal and spatial scales and in quantifying the associated model parameters. This in itself has led to an entire branch of the field dedicated to model parameterization, parameter estimation, and model calibration. Many model intercomparison projects (MIPs) over the past decades have tried to determine which models or model parameterizations perform best under certain conditions and which provide a better description of the underlying processes. With few exceptions, they have had limited success for a number of reasons. Differences in model parameters can compensate for differences in model parameterizations; most MIPs only evaluate model differences for a narrow range of conditions; and model evaluation metrics tell us about model performance, but fail to provide insight into model behavior.

One avenue being explored to make this model diversity more accessible and to explore differences in model behavior is based on concepts of models with flexible model structures. The Structure for Unifying Multiple Modeling Alternatives (SUMMA) is one such modeling framework. SUMMA provides flexible representations for spatial arrangements, flux parameterizations, and parameter values. This makes it an ideal modeling framework for investigating the effects of different model structures. In addition, new methods for model evaluation are finding their way into hydrology. One promising avenue is the use of information theory to evaluate model results. Ideally, this would provide insight into model behavior rather than yet another measure of model performance. If so, then we can use this knowledge to guide model selection based on available information.

In this work, we instantiate a large number of SUMMA runs with different modeling options across sites in diverse hydrometeorological regimes and analyze the full set of energy and mass balance terms to quantify the effects that model structure has on model output. To do so, we calculate information theoretic process networks from the time series outputs of the model instantiations. These high-dimensional networks are then analyzed via clustering and dimensionality reduction algorithms to develop model "fingerprints" that provide high-level insights into the model structure space that is covered.