Predicting the lumber volume recovery of Picea mariana using parametric and non-parametric regression methods

Three model forms (polynomial, power and exponential) and a local regression (LOESS) model with different variables were studied for their goodness of predicting lumber volume recovery from two types of sawmill (stud mill and optimized random mill). Explanatory variables used to predict lumber volume recovery were three basic tree characteristics: diameter at breast height (dbh), tree height and tree taper. Based on the selected statistical criteria such as R-2 mean absolute prediction error (MAE) and root mean square error (RMSE), the polynomial functions, power functions and the local regression (LOESS) model, in general, had excellent abilities to predict lumber volume recovery. The simplified second order polynomial model with both dbh and tree height variables predicted the lumber volume recovery almost as accurately as did the LOESS model with the same predictor variables. Model validation using independent data from a real stud mill indicated that the two model forms were able to forecast lumber volume recovery from measured tree characteristics, especially for small and medium-sized trees.