Modeling long-term forest dynamics under global change: The trade-off between generality and (local) accuracy

Background/Question/Methods

Since the 1980s, dynamic models are being used to assess the fate of forest structure and composition under the impacts of global change drivers. While early applications focused on the general trends of forest properties through time, over the last decade the emphasis has moved to providing locally accurate projections aimed at supporting local to regional decision-making. This has necessitated more detail to be added to the models, and the question arises to which extent the models are still applicable under changing climatic conditions, or whether they have become restricted to local (current) conditions. We report on recent experiences in this context based on the development and application of the ForClim model, which at one time was a model that was broadly applicable to widely different areas such as the Pacific Northwest, central Europe, and northeastern China.

Results/Conclusions

Recent developments of ForClim have led to high local accuracy of the model as measured e.g. by its ability to track time series of forest inventory data (basal area, tree numbers) at climatically very different sites in Switzerland, extending over as much as 105 years. We also found that this model version has reasonable performance when applied along a transect from the Pacific coast to the interior of Oregon, suggesting that it is possible to achieve high local accuracy while maintaining model generality. Yet, model performance deteriorated dramatically at sites nearby those for which the model was refined, e.g. in Slovenia, which are characterized by highly similar environmental conditions and tree species as in Switzerland. The conclusions from these exercises are twofold: first, it is necessary to better understand local adaptations of tree species to the prevailing climatic conditions (ecotypic variation, phenotypic plasticity); and second, we need to be careful not to overfit ecological models so as not to constrain their wide applicability.