A Bayesian approach for calibrating tree mortality in a forest succession model

Background/Question/Methods

The applicability of forest succession models to a wide range of sites and tree species under various climatic conditions makes them promising tools to predict potential shifts of forest structure and composition due to climate change. However, most of these models are still missing an empirically-based representation of a fundamental ecological process: tree mortality. Latest research about tree growth-mortality relationships has provided forest succession modelers with more options to replace former, simple mortality algorithms. However, such approaches, although ecologically sound, may not lead automatically to model improvement as parameterization is often based on limited data. Moreover, simplified growth-mortality interactions within succession models could make a recalibration necessary. A promising approach for such a recalibration is to use Bayesian techniques, where latest developments have widened the field of application to parameter-rich process-based models. In this study, we implemented an inventory-based mortality routine (IBM) into the forest succession model ForClim and simulated potential natural vegetation (PNV) along an environmental gradient across Switzerland. We applied approximate Bayesian techniques to recalibrate the routine’s parameters using data from natural forest reserves, and compared posterior parameter estimates to prior ones. Finally, we repeated PNV simulations to test the performance of this recalibrated model version.

Results/Conclusions

The IBM model version (without recalibration) reproduced basal area and stem numbers in acceptable magnitude. However, it totally failed to portray the diversity of tree species composition, which should be expected along such an environmentally heterogeneous gradient. Shade-tolerant species were over-represented, while drought- and cold- tolerant species were neglected. Parameter estimates of the IBM strongly changed when it was fitted in order that ForClim simulations matched forest inventory data. In general, the systematic recalibration using Bayesian techniques lead to a considerable improvement of model performance and plausibility at most studied sites. Furthermore, this approach was supportive in identifying structural problems in the growth and establishment sub-models of ForClim:For instance, Bayesian fitted parameters of IBM were based on simulated growth rates while the original parameters were calculated based on real growth data. In consequence, they may accurately reflect process interactions within the model but fail in terms of plausibility (posterior distribution not realistic). By using Bayesian techniques, such mismatches between accuracy and plausibility in model algorithms can be quantitatively assessed, which is a key step for further model improvement.