Calibrating a process-based forest model with a rich observational dataset at 22 European forest sites

Abstract. In recent years model-data interaction has improved through use of probabilistic techniques to inform and reduce the uncertainty of model parameters, while also taking into account observational uncertainty. This study builds on previous work, through access to a richer representation of the plant-soil ecosystem at multiple European forest sites, than was previously available. Given this rich dataset, we asked which observational datasets were most effective in reducing uncertainty in model predictions and model-data differences. Also, since there is a lack of consensus about whether it is more beneficial to calibrate forest sites separately or together we revisited this question with a particular emphasis on which is most effective in reducing model-data differences and uncertainty. We performed single dataset Bayesian calibrations (BC) and compared the results with a calibration with all the observations included. We also compared calibrations where each pine forest site was calibrated separately with a calibration where all the pine sites were calibrated together. While measurements of plant and soil carbon stocks were more sparse, their inclusion in the BC were more important for reducing model-data differences and uncertainty in the above and belowground carbon pools than the greater numbers of carbon and water flux data. Our results suggest that use of calibration data representing just a few aspects of the ecosystem could be problematic, since improved model-data fits for the parts of the system represented by the data could be at the expense of other part of the system, where the model-data fit worsened. The single dataset calibrations helped to diagnose where there may be inconsistencies between different datasets or between the model and data or both. These inconsistencies hampered the reduction in model-data differences in the calibration with all the observations present. As expected, we found a strong relationship between the quantity of data included in the calibration and the uncertainty reduction after BC, finding the largest reduction in uncertainty when all the observations were included. For some ecosystem variables uncertainty reduced after calibration but model-data differences increased. This would suggest that there were deficiencies in the model or systematic errors in the data or both. These results advocate the use of calibration datasets which represent the rich diversity of the ecosystem under investigation but where model discrepancies and data systematic errors are explicitly represented in the BC. While separate calibrations at each forest site generally reduced model-data differences more than calibrating at all the sites together, parts of the ecosystem that were sparsely observed benefited more from the multi-site calibration. Multi-site calibration led to larger and more consistent reductions in uncertainty than separate calibrations at each site, especially for ecosystem variables with fewer observations. These results support the use of Bayesian hierarchical calibration which allows variation in model parameters between different sites while allowing information to be shared across sites for sparsely observed ecosystem variables.