Leaf and wood classification framework for terrestrial LiDAR point clouds

Leaf and wood separation is a key step to allow a new range of estimates from Terrestrial LiDAR data, such as quantifying above-ground biomass, leaf and wood area and their 3D spatial distributions. We present a new method to separate leaf and wood from single tree point clouds automatically. Our approach combines unsupervised classification of geometric features and shortest path analysis. The automated separation algorithm and its intermediate steps are presented and validated. Validation consisted of using a testing framework with synthetic point clouds, simulated using ray-tracing and 3D tree models and 10 field scanned tree point clouds. To evaluate results we calculated accuracy, kappa coefficient and F-score. Validation using simulated data resulted in an overall accuracy of 0.83, ranging from 0.71 to 0.94. Per tree average accuracy from synthetic data ranged from 0.77 to 0.89. Field data results presented and overall average accuracy of 0.89. Analysis of each step showed accuracy ranging from 0.75 to 0.98. F-scores from both simulated and field data were similar, with scores from leaf usually higher than for wood. Our separation method showed results similar to others in literature, albeit from a completely automated workflow. Analysis of each separation step suggests that the addition of path analysis improved the robustness of our algorithm. Accuracy can be improved with per tree parameter optimization. The library containing our separation script can be easily installed and applied to single tree point cloud. Average processing times are below 10min for each tree.