Analysis and comparison of nonlinear tree height prediction strategies for Douglas-fir forests

Using an extensive Douglas-fir data set from southwest Oregon, we examined the (1) performance and suitability of selected prediction strategies, (2) contribution of relative position and stand-density measures in improving tree height (h) prediction values, and (3) effect of different subsampling designs to fill in missing h values in a new stand using a regional nonlinear model. Nonlinear mixed-effects models (NMEM) substantially improved the accuracy and precision of height prediction over the conventional nonlinear fixed-effects model (NFEM) that assumes the observations are independent, particularly when a few trees are subsampled for height. The predictive performance of a correction factor on a NFEM with relative position and stand-density measures was comparable to that of a NMEM when four or more trees were subsampled for height. When two or more heights were randomly subsampled, the NMEM efficiently explained the differences in the height-diameter relationship because of the variations in relative position of trees and stand density without having to incorporate them into the model. When only one height was subsampled, selecting the largest diameter tree in the stand would result in a lower predicted root mean square error (RMSE) than randomly selecting the height, regardless of the model form or fitting strategy used.