Predicting site index of plantation loblolly pine from biophysical variables

Concerns of the effect of climate change on forest productivity have impelled the need to accurately predict forest productivity from climate, physiographic and edaphic variables (biophysical variables). We fitted and evaluated random forest models and nonlinear least squares regression models for predicting plantation loblolly pine (Pinus taeda L.) site index from biophysical variables. Tree and stand location data were provided by the Virginia Tech Forest Modeling Research Cooperative. Climate data for each stand location were computed using the Oakridge National Laboratories' daily surface weather prediction models, while soils data were extracted from the USDA Natural Resource Conservation Service SSURGO GIS database using GIS data extraction techniques. Separate models were fitted for non-intensively managed (Non-IMP) and intensively managed (IMP) loblolly pine plantations. Variable selection methods in both modeling approaches showed that the number of biophysical variables that were important in predicting site index of IMP loblolly pine was smaller than the number for Non-IMP stands. The non-parametric random forest models had better fit and prediction statistics than the least squares parametric models but exhibited the potential to give illogical predictions under extrapolation. Site index predictions from both modeling approaches exhibited a regression towards the mean.