A comparison between traditional ordinary least-squares regression and three methods for enforcing additivity in biomass equations using a sample of Pinus radiata trees
Mohan, K. C.
Mason, Euan G.
Bown Intveen, Horacio
Cita de ítem
New Zealand Journal of Forestry Science (2020) 50:7
Background: Additivity has long been recognised as a desirable property of systems of equations to predict the biomass of components and the whole tree. However, most tree biomass studies report biomass equations fitted using traditional ordinary least-squares regression. Therefore, we aimed to develop models to estimate components, subtotals and above-ground total biomass for a Pinus radiata D.Don biomass dataset using traditional linear and nonlinear ordinary leastsquares regressions, and to contrast these equations with the additive procedures of biomass estimation.
Methods: A total of 24 ten-year-old trees were felled to assess above-ground biomass. Two broad procedures were implemented for biomass modelling: (a) independent; and (b) additive. For the independent procedure, traditional linear models (LINOLS) with scaled power transformations and y-intercepts and nonlinear power models (NLINOLS) without y-intercepts were compared. The best linear (transformed) models from the independent procedure were further tested in three different additive structures (LINADD1, LINADD2, and LINADD3). All models were evaluated using goodness-of-fit statistics, standard errors of estimates, and residual plots.
Results: The LINOLS with scaled power transformations and y-intercepts performed better for all components, subtotals and total above-ground biomass in contrast to NLINOLS that lacked y-intercepts. The additive model (LINADD3) in a joint generalised linear least-squares regression, also called seemingly unrelated regression (SUR), provided the best goodness-of-fit statistics and residual plots for four out of six components (stem, branch, new foliage and old foliage), two out of three subtotals (foliage and crown), and above-ground total biomass compared to other methods. However, bark, cone and bole biomass were better predicted by the LINOLS method.
Conclusions: SUR was the best method to predict biomass for the 24-tree dataset because it provided the best goodness-of-fit statistics with unbiased estimates for 7 out of 10 biomass components. This study may assist silviculturists and forest managers to overcome one of the main problems when using biomass equations fitted independently for each tree component, which is that the sum of the biomasses of the predicted tree components does not necessarily add to the total biomass, as the additive biomass models do.