Prediction of diameter distributions and tree-lists in southwestern Oregon using LiDAR and stand-level auxiliary information

Diameter distributions and tree-lists provide information about forest stocks disaggregated by size and species and are key for informing forest management. Diameter distributions and tree-lists are multivariate responses, which makes the evaluation of methods for their prediction reliant on the use of dissimilarity metrics to summarize differences between observations and predictions. We compared four strategies for selection of k nearest neighbors (k-NN) methods to predict diameter distributions and tree-lists using LiDAR and stand-level auxiliary data and analyzed the effect of the k-NN distance and number of neighbors in the performance of the predictions. Strategies differed by the dissimilarity metric used to search for optimal k-NN configurations and the presence or absence of post-stratification. We also analyzed how alternative k-NN configurations ranked when tree-lists were aggregated using different DBH classes and species groupings. For all dissimilarity metrics, k-NN configurations using random-forest distance and three or more neighbors provided the best results. Rankings of k-NN configurations based on different dissimilarity metrics were relatively insensitive to changes on the width of the DBH classes and the definition of the species groups. The selection of the k-NN methods was clearly dependent on the choice of the dissimilarity metric. Further research is needed to find suitable ways to define dissimilarity metrics that reflect how forest managers evaluate differences between predicted and observed tree-lists and diameter distributions.