Using tree clusters to derive forest properties from small footprint lidar data

This paper describes a new object-oriented small footprint lidar algorithm in which the objects of interest are tree clusters. The algorithm first thresholds the lidar canopy height model (CHM) at two levels to produce tree cluster grids. Next, two metrics are calculated based on these grids. The metric values are used in a multiple regression equation to predict the forest parameter of interest. To set the two thresholds, an optimization algorithm is used in conjunction with training data consisting of subsets of the CHM in which the forest parameters are known through ground measurements. A test of the algorithm was performed using ground and lidar data from a non-intensively managed loblolly pine (Pinus taeda) plantation in Virginia. The accuracies of the lidar-based predictions of density (0.01 <= 5 R-2 <= 0.80; 126 trees/ha <= RMSE <= 8,173 trees/ha) and biomass (0.04 <= R-2 <= 0.62; 12.4 t/ha <= 316.5 t/ha) depended on the combination of metrics used, whether trees with a diameter at breast height < 10 cm were excluded from the analysis, and the number of plots used for training and testing. However, the fit between the ground measurements and tree cluster-based predictions generally exceeded the fit between ground measurements and the output from an individual tree-based algorithm tested using the same data (100 percent of comparable cases when density was predicted, 85 percent of comparable cases when biomass was predicted, based on the coefficient of determination and RMSE).