A comparison of tree selection strategies for developing allometric biomass models in the context of model application

Abstract

Improving forest biomass and carbon estimates is essential for sustaining the mitigation of climate change efforts in the forestry sector. Allometric biomass models, which relate biomass to easily measured tree attributes such as diameter at breast height (D) and height, are the primary tool used to estimate forest carbon stocks. When developing allometric biomass models, the process of selecting trees to calibrate the model is an important step that affects the model’s predictive performance. Typically, the D-range of the species is divided into D-classes, followed by random selection of sample trees from within the D-classes; the number of trees in each D-class defines the sample tree D-distribution. Here, using a simulation study, we compared eight types of sample tree D-distributions with respect to the precision of estimates of mean population biomass per hectare that the models produced. The results showed that, for a given sample size, when the proportion of the sample to be randomly selected from each D-class matches the proportion of the basal area in each D-class in the population, the D-distribution was optimal with respect to minimizing the standard error of the estimate of the population mean per hectare. The Neyman allocation approach and the D-distribution based on biomass instead of basal area produced very similar results to those of the optimal strategy. When a-priori information about the distribution of tree attributes in the population is unknown, a uniform D-distribution represents a good alternative to the optimal sample D-distribution. Although producing the greatest precision of the estimate, the optimal D-distribution of sample trees did not produce models with the most accurate model fit (greatest coefficient of determination), suggesting that quality of fit metrics alone are not sufficient indicators to evaluate the effects of model prediction uncertainty on the precision of a subsequent population parameter estimate. The sample tree D-distribution had considerable effects on the precision of estimates of mean population biomass that the models produced. To optimize the tree selection, and, therefore, to develop the models that for a given sample size produce precise estimators of the population mean, we recommend using an optimal D-distribution of the sample trees.