Process-Level Diagnostics of Terrestrial Hydrology Models

To make reliable predictions under nonstationary conditions we must build physically realistic models. However, because hydrological systems are compendiums of dynamic interactions between many different components and processes, it is difficult to use a direct comparison between model predictions and observations of integrated system responses to determine exactly which process representation(s) in any given model might contribute to or compensate for prediction error. Hydrologists and Earth Systems Modelers might recognize this problem as one of model diagnostics, however this is a classical problem that affects almost all aspects of scientific investigation – it is the confirmation holism aspect of the Duhem-Quine thesis.

We propose here a systematic way to attack this problem that is based on a fundamental logic-based interpretation of the Duhem-Quine problem. The general idea is that diagnostic evaluation of complex systems models will involve tracking information flows through and between different interacting components in a given model structure, and that it is actually these information flows that we should wish to validate, evaluate or benchmark against observations. The problem is that we rarely have observations of all pertinent states and fluxes at all relevant spatiotemporal scales, and we propose that the fundamental resolution to this problem is data assimilation. The key insight is that data assimilation is simply the projection of information onto the states of a dynamical systems model. We discuss the implications of this for doing science and making predictions with coupled land surface hydrology models, as well as the risks associated with using sub-optimal data assimilation strategies.

We will finally outline a few application examples where we found that land surface models apparently have a systematic problem with underestimating the connectivity between hydrological and ecological processes. We will us these examples to show how data-assimilation-based process-level model evaluation can lead to better insights about the reliability and realism of complex terrestrial hydrology models than can response-based model evaluation.