Aggregating Pixel-Level Basal Area Predictions Derived from LiDAR Data to Industrial Forest Stands in North-Central Idaho

Stand exams are the principal means by which timber companies monitor and manage their forested lands. Airborne LiDAR surveys sample forest stands at much finer spatial resolution and broader spatial extent than is practical on the ground. In this paper, we developed models that leverage spatially intensive and extensive LiDAR data and a stratified random sample of field plots across two mixed conifer forest landscapes in north-central Idaho. Our objective was to compare alternative models for producing unbiased maps of basal area per acre (BAA; ft2/acre), towards the greater goal of developing more accurate and efficient inventory techniques. We generated 60 topographic or stand structure metrics from LiDAR that were used as candidate predictor variables for modeling and mapping BAA at the scale of 30m pixels. Tree diameters were tallied in 1/10 and 1/5 acre fixed-radius plots (N = 165). Four models are presented, all based on 12 predictor variables. The first imputes BAA as an auxiliary variable from an imputation model that uses the machine learning algorithm randomForest in classification mode, and was developed in a prior study to map species-level basal areas of 11 conifer species; the second uses randomForest in regression mode to predict BAA as a single response variable from these same 12 predictor variables. The third is a linear regression model that predicts ln-transformed BAA using a best subset of 12 different predictor variables; the fourth again uses randomForest in regression mode, based on the same best subset of 12 variables selected for the linear regression model. We aggregated the pixel-level predictions within industrial forest stand boundaries, and then used equivalence plots to evaluate how well the aggregated predictions matched independent stand exams (having projected the tree growth in FVS and updated the stand tables to July 2003, the time of the LiDAR acquisition). All four models overpredicted BAA, but the bias was significant only in the case of the regression model. Predictions from the two randomForest models run in regression mode were very similar, despite using different predictor variables. We conclude that randomForest can be used to impute or predict canopy structure information from LiDAR-derived topographic and structural metrics with sufficient accuracy for operational management of conifer forests. In the future, tree lists could be imputed from LiDAR-derived canopy structure metrics empirically related to plot-level tree measurements. This will allow projections of tree growth at the pixel level across forested landscapes, instead of at the stand level as is the current norm.