Estimating model- and sampling-related uncertainty in large-area growth predictions

Estimating uncertainty in forest growth predictions is essential to support large-area policies and decisions. The aim of this study was to estimate model and sampling uncertainties at a regional level. To do this, we generated forest growth predictions for three ecotypes in the Bas-Saint-Laurent region of Quebec, Canada. Predictions were generated using the ARTEMIS growth model that allows for stochasticity in some of the sub-models. We used a bootstrap hybrid estimator to estimate the variances arising from the model and the sampling. Moreover, the variance due to the model was further decomposed to determine which dynamic sub-model induced the greatest share of variance. Results revealed that sampling accounted for most of the variance in short-term predictions, In long-term predictions, the model contribution turned out to be as important as that of the sampling. The variance decomposition per sub-model indicated that the mortality sub-model induced the highest variability in the predictions. These results were consistent for the three ecotypes. We recommend that efforts in variance reduction focus on increasing the sample size in short-term predictions and on improving the mortality sub-model in long-term predictions.