Modelling canopy gap probability, foliage projective cover and crown projective cover from airborne lidar metrics in Australian forests and woodlands

Tree canopy density metrics (TCDM) derived from airborne lidar data are used in a range of crucial environmental monitoring, forestry and natural resource management applications. The derivation of spatially and temporally consistent TCDM, however, typically requires field calibration to account for differences in instrument/survey parameters. Lidar surveys with no coincident field measurements consequently will have an unknown error associated with TCDM limiting their application. In this study, we analysed an extensive set of lidar captures with coincident field data to determine the lidar TCDM that best match the canopy gap probability (P-gap), foliage projective cover (FPC) and crown projective cover (CPC). Furthermore, we developed and evaluated models designed to reduce the bias introduced by variations in lidar instrument and survey acquisition parameters. The dataset incorporated 148 field sites (100 m diameter circular plots) coincident with 13 different lidar surveys between 2008 and 2015, distributed across a range of Australian forests and woodlands. The best lidar metric for 1 - P-gap, achieving a root mean square error (RMSE) of 6.7% with 95% confidence intervals (CI) of 6.1-7.3%, was the proportion of all returns greater than a canopy height threshold (tcanopy) of 1.5 m above ground (dall). The best metric for FPC (RMSE = 6.0%, CI = 5.3-6.7%) used the proportion of returns, weighted as the fraction of the number of returns recorded from each pulse (dweighted), with tcanopy of 1.7 m. The best metric for CPC (RMSE = 7.0%, CI = 6.4-7.7%) was the proportion of 0.5 m pixels greater than 0.8 m above the ground, for an interpolated canopy height model (dinterp). Overall bias for these metrics was low (similar to 1%), however, the bias for individual surveys varied significantly. For example, for one survey dall consistently underestimated 1 - P-gap with a bias of -8.3%, while a different survey consistently overestimated 1 - P-gap with a bias of 3.8%. Elastic net regression models, using instrument, survey and plot parameters as predictor variables, were unable to consistently remove the bias. No relationships could be discerned between lidar parameters and the bias between lidar metrics and field measurements, potentially due to complex interactions between parameters, the spatial scale of the field plots, and uncertainties in field measurements and lidar attributes. Although the bias could not be modelled, the results provide metrics to derive P-gap, FPC and CPC with less than 10% error from lidar surveys captured with similar parameters across Australia (>600,000 km(2)).