Individual tree detection and crown delineation from unmanned aircraft system (UAS) LiDAR in structurally complex mixed species eucalypt forests

Estimates of forest stocking density per hectare (NHa) are important in characterising ecological conditions and assessing changes in forest dynamics after disturbances due to pyrogenic, anthropogenic and biotic factors. We use Unmanned Aircraft Systems (UAS) LiDAR with mean point density of 1485 points m−2 across 39 flight sites to develop a bottom-up approach for individual tree and crown delineation (ITCD). The ITCD algorithm was evaluated across mixed species eucalypt forests (MSEF) using 2790 field measured stem locations across a broad range of dominant eucalypt species with randomly leaning trunks and highly irregular intertwined canopy structure. Two top performing ITCD algorithms in benchmarking studies resulted in poor performance when optimised to our plot data (mean Fscore: 0.61 and 0.62), which emphasises the challenge posed for ITCD in the structurally complex conditions of MSEF. To address this, our novel bottom-up ITCD algorithm uses kernel densities to stratify the vegetation profile and differentiate understorey from the rest of the vegetation. For vegetation above understorey, the ITCD algorithm adopted a novel watershed clustering procedure on point density measures within horizontal slices. A Principal Component Analysis (PCA) procedure was then applied to merge the slice-specific clusters into trunks, branches, and canopy clumps, before a voxel connectivity procedure clustered these biomass segments into overstorey trees. The segmentation process only requires two parameters to be calibrated to site-specific conditions across 39 MSEF sites using a Shuffled Complex Evolution (SCE) optimiser. Across the 39 field sites, the ITCD algorithm had mean Fscore of 0.91, True Positive (TP) trees represented 85% of measured trees and predicted plot-level stocking (NP) averaged 94% of actual stocking (NOb). As a representation of plot-level basal area (BA), TP trees represented 87% of BA, omitted trees represented slightly smaller trees and made up 8% of BA, and a further 5% of BA had commission error. Spatial maps of NHa using 0.5 m grid-cells showed that omitted trees were more prevalent in high density forest stands, and that 63% of grid-cells had a perfect estimate of NHa, whereas a further 31% of the grid-cells overestimate or underestimate one tree within the search window. The parsimonious modelling framework allows for the two calibrated site-specific parameters to be predicted (R2: 0.87 and 0.66) using structural characteristics of vegetation clusters within sites. Using predictions of these two site-specific parameters across all sites results in mean FScore of 0.86 and mean TP of 0.77, under circumstances where no ground observations were required for calibration. This approach generalises the algorithm across new UAS LiDAR data without undertaking time-consuming ground measurements within tall eucalypt forests with complex vegetation structure