How Does Higher Frequency Monitoring Data Affect the Calibration of a Process-Based Water Quality Model?

Process-based catchment water quality models are increasingly used as tools to inform land management. However, for such models to be reliable they need to be well calibrated and shown to reproduce key catchment processes. Calibration can be challenging for process-based models, which tend to be complex and highly parameterised. Calibrating a large number of parameters generally requires a large amount of monitoring data, but even in well-studied catchments, streams are often only sampled at a fortnightly or monthly frequency. The primary aim of this study was therefore to investigate how the quality and uncertainty of model simulations produced by one process-based catchment model, INCA-P (the INtegrated CAtchment model of Phosphorus dynamics), were improved by calibration to higher frequency water chemistry data. Two model calibrations were carried out for a small rural Scottish catchment: one using 18 months of daily total dissolved phosphorus (TDP) concentration data, another using a fortnightly dataset derived from the daily data. To aid comparability, calibrations were carried out automatically using the MCMC-DREAM algorithm.

Using daily rather than fortnightly data resulted in improved simulation of the magnitude of peak TDP concentrations, in turn resulting in improved model performance statistics. Marginal posteriors were better constrained by the higher frequency data, resulting in a large reduction in parameter-related uncertainty in simulated TDP (the 95% credible interval decreased from 26 to 6 μg/l). The number of parameters that could be reliably auto-calibrated was lower for the fortnightly data, leading to the recommendation that parameters should not be varied spatially for models such as INCA-P unless there is solid evidence that this is appropriate, or there is a real need to do so for the model to fulfil its purpose.

Secondary study aims were to highlight the subjective elements involved in auto-calibration and suggest practical improvements that could make models such as INCA-P more suited to auto-calibration and uncertainty analyses. Two key improvements include model simplification, so that all model parameters can be included in an analysis of this kind, and better documenting of recommended ranges for each parameter, to help in choosing sensible priors.