On the dialog between experimentalist and modeler in catchment hydrology: Use of soft data for multicriteria model calibration

Abstract

The dialog between experimentalist and modeler in catchment hydrology has been minimal to date. The experimentalist often has a highly detailed yet highly qualitative understanding of dominant runoff processes; thus there is often much more information content on the catchment than we use for calibration of a model. While modelers often appreciate the need for “hard data” for the model calibration process, there has been little thought given to how modelers might access this “soft” or process knowledge. We present a new method where soft data (i.e., qualitative knowledge from the experimentalist that cannot be used directly as exact numbers) are made useful through fuzzy measures of model simulation and parameter value acceptability. We developed a three‐box lumped conceptual model for the Maimai catchment in New Zealand, a particularly well‐studied process‐hydrological research catchment. The boxes represent the key hydrological reservoirs that are known to have distinct groundwater dynamics, isotopic composition, and solute chemistry. The model was calibrated against hard data (runoff and groundwater levels) as well as a number of criteria derived from the soft data (e.g., percent new water, reservoir volume, etc.). We achieved very good fits for the three‐box model when optimizing the parameter values with only runoff (R eff = 0.93). However, parameter sets obtained in this way showed in general a poor goodness of fit for other criteria such as the simulated new water contributions to peak runoff. Inclusion of soft data criteria in the model calibration process resulted in lower R eff values (around 0.84 when including all criteria) but led to better overall performance, as interpreted by the experimentalist's view of catchment runoff dynamics. The model performance with respect to soft data (like, for instance, the new water ratio) increased significantly, and parameter uncertainty was reduced by 60% on average with the introduction of the soft data multicriteria calibration. We argue that accepting lower model efficiencies for runoff is “worth it” if one can develop a more “real” model of catchment behavior. The use of soft data is an approach to formalize this exchange between experimentalist and modeler and to more fully utilize the information content from experimental catchments.