Individual tree segmentation from a leaf-off photogrammetric point cloud

The proliferation of small unmanned aerial vehicles and structure from motion photogrammetry provides simple, repeatable, and inexpensive techniques for creating high-resolution aerial imagery and 3D data. These methods have been quickly adopted within the forestry realm for mapping forest canopies, among other tasks. Here, we present a workflow to capture sub-canopy structure in a leaf-off deciduous forest with a hobby-level drone and geographical information system software, and a toolkit for segmenting and measuring individual tree diameter from the resultant point cloud. Our methods, which build on steps from several previously published approaches (layer stacking for identifying vertical features and iterative neighborhood searches), rely solely on structural characteristics and could therefore be applied to dense photogrammetric and laser-scanning point clouds. Across four variable study sites, the toolkit correctly identified 279 trees (recall=0.89, precision=0.97, F-score=0.93) and correctly segmented 90.9% of points in the point cloud (r=0.98, p=0.93, F=0.95). Point cloud measures of diameter at breast height generally correlated with field measurements (Pearson's r=0.82, root mean square error=0.11m), but were typically higher than field measurements. These results demonstrate the capacity for assessing forest structure and conducting inventories using widely accessible hardware and software.