Do higher data frequency and Bayesian auto-calibration lead to better model calibration? Insights from an application of INCA-P, a process-based river phosphorus model

We use Bayesian auto-calibration to explore how observed data frequency affects the performance and uncertainty of INCA-P, a process-based catchment phosphorus model. A fortnightly dataset of total dissolved phosphorus (TDP) concentration was derived from 18 months of daily data from a small (51 km(2)) rural catchment in northeast Scotland. We then use the DiffeRential Evolution Adaptive Metropolis (DREAM) Markov Chain Monte Carlo (MCMC) algorithm to calibrate 29 of the >127 model parameters using the daily and the fortnightly observed datasets. Using daily rather than fortnightly data for model calibration resulted in a large reduction in parameter-related uncertainty in model output and lower risk of obtaining unrealistic results. However, peaks in TOP concentration were as well simulated as when fortnightly data were used. A manual model calibration did a better job of simulating the magnitude of TDP peaks and baseflow concentrations, suggesting that alternative measures of model performance may be needed in the auto-calibration. Results suggest that higher frequency sampling, perhaps for just a short period, can greatly increase the confidence that can be placed in model output. In addition, we highlight the many subjective elements involved in auto-calibration, in an attempt to temper a common perception that auto-calibration is an objective and rigorous alternative to manual calibration. Finally, we suggest practical improvements that could make models such as INCA-P more suited to auto-calibration and uncertainty analyses, a key requirement being model simplification. (C) 2015 Elsevier B.V. All rights reserved.