Comparison of models describing forest inventory attributes using standard and voxel-based lidar predictors across a range of pulse densities

Abstract Fine-scale characterisation of forest stands using very high-density aerial lidar data holds considerable potential for improving the accuracy of area-based forest inventories. To realise these gains, new methods of characterising dense aerial point clouds are required. This research presents one potential approach using voxel-based metrics often associated with the analysis of terrestrial lidar data. This was accomplished by comparing predictions of forest inventory attributes made using voxel-based metrics, more standard lidar metrics and a combination of both classes of metrics. A high-density lidar dataset was acquired using a helicopter-mounted RIEGL VUX-1UAV laser scanner to produce point clouds with a minimum density of 280 pulses m −2 . Data were obtained from 73 plots presenting a wide range of stand conditions located within two adjacent plantations of Pinus radiata D.Don in south-eastern New South Wales. Random forests regression models were developed to predict top height, basal area, stand density and total stem volume. To assess the interaction between metric type and pulse density, the point clouds were thinned to 18 pulse densities ranging from 1 to 280 pulses m −2 before fitting models using the metrics generated from data at each target density. Data thinning had little effect on the predictive accuracy of models for any of the four forest attributes predicted from either voxel-based, standard lidar metrics or their combination. Averaged across all pulse densities, models created for top height, basal area, stand density and total stem volume from standard lidar metrics had R 2 of 0.72, 0.44, 0.34 and 0.53 with normalised RMSE (RMSE expressed as a percentage of the mean for each dimension) of 6,6, 25.2, 60.1 and 25.5% respectively. Use of voxel-based metrics resulted in substantial gains in model precision for all dimensions, apart from top height, with R 2 increasing by 0.04, 0.23, 0.24, and 0.22 and nRMSE averaging 6.1, 19.6, 48.6, and 18.7%, respectively, for top height, basal area, stand density, and total stem volume. The precision of models that used both types of lidar metrics was very similar to the precision of models that used only voxel-based metrics. These results demonstrate the considerable potential of voxel-based metrics for improving the accuracy of forest measurement. The gains from voxelised-metrics were not dependent on very high pulse densities and could be achieved at densities typical of conventional lidar surveys undertaken using fixed-wing aircraft.