Determining an optimal model for processing lidar data at the plot level: results for a Pinus radiata plantation in New South Wales, Australia

Small-footprint, discrete return airborne laser scanning (ALS or lidar) data is increasingly being used by forest managers to assist forest inventories. In this study, airborne lidar and plot-based data were collected from a 5 000 ha study site within Green Hills State Forest, a Pinus radiata D.Don plantation in southern New South Wales, Australia. A series of area-based lidar metrics were extracted and modelled against four inventory attributes (mean tree height, stem density, basal area and stand volume) obtained from 63 ground plots. For all response variables, regression tree models had the best model fit compared to Random Forest and Bayesian Model Averaging modelling techniques. The best regression tree models were based on the lidar metrics: the 5th and 95th height percentiles, minimum vegetation height, density of non-ground returns and a measure of spatial variation, the rumple index. All these metrics can be easily derived from the lidar data. The best regression tree models for each inventory attribute produced the following R2 values: for mean tree height (m), R2 = 0.94; stocking (trees ha-1), R2 = 0.85; basal area (m2 ha-1), R2 = 0.81 and for stand volume, R2 = 0.81 (m3 ha-1) while the corresponding relative RMSEs were 5.8%, 23.4%, 15.5% and 22.3%, respectively. These models were then used to produce prediction maps over a 50 m grid across the 5 000 ha study site. Results from this study support the operational inclusion of airborne lidar data within P. radiata resource inventory systems.