Parametric vs. nonparametric LiDAR models for operational forest inventory in boreal Ontario

Parametric and nonparametric predictions of forest inventory attributes from airborne LiDAR data are compared for a forest management unit in boreal Ontario. For the parametric approach, seemingly unrelated regression models were calibrated by forest type (SUR) and for all forest types combined (SUR_All). For the nonparametric approach, randomForest (RF) and k-nearest neighbours (kNN) were implemented. Calibration data consisted of 442 circular 0.04 ha plots covering a range of development stages within eight forest types. Results were validated on 64 independent plots distributed across the same forest types. Predicted variables included top height, merchantable basal area, and gross merchantable volume. In general, RF and SUR predictions were the most accurate and precise, whereas kNN and SUR_All predictions were less reliable. Prediction accuracy and precision varied markedly with forest type, with no single method producing results that were consistently best. None of the methods extrapolated well, underscoring the need to capture the full range of population variation during calibration. Parametric predictions were improved by forest-type stratification, necessitating a population forest-type layer prior to application. In contrast, forest type was not an important predictor in the nonparametric solutions. RF can offer significant operational advantages over parametric regression without loss of accuracy or precision.