Forests should be monitored and measured frequently to ensure its sustainability and continuity of carbon sequestration. However, conventional monitoring methods are costly and labor intensive. New technologies, such as unmanned aircraft systems (UAS) can be used as an alternative method to monitor and measure forests. The objective of this study is to evaluate the impact of UAS-based data processing algorithms on forest structure estimation. The data for this study was collected with a Sony A5100 mounted on a Tarot 650 quadcopter flown over a Douglas-fir plantation forest near Lincoln City, Oregon, USA.
3D point clouds were produced from the images acquired with two commercial software packages: Agisoft and Pix4D. The point clouds were generated using two sets of parameters, each with two different values. For Agisoft, the parameters were image alignment (i.e., high and medium) and dense point quality (i.e., high and medium). For Pix4D, the parameters were image scale (i.e., original and half) and point density (i.e., optimal and low). Individual trees were delineated from the point clouds using three tree segmentation algorithms: variable window filter (implemented in ForestTools Package in R), Graph-Theoretical (implemented in TrEx software), and Watershed segmentation (implemented in ArcMap). The quality of the segmentation was assessed by comparing the computed trees with the actual trees manually identified on nine 0.05 ha plots systematically distributed throughout the study area.
Results showed that different data processing and tree segmentation algorithms led to significantly different forest structure estimations. On average, the tree segmentation algorithms used in this study had relatively similar percent number of correctly matched trees when applied with Agisoft generated point clouds and strong consistency with correctly matched percentages between 83% and 84.61%. With the Pix4D generated point clouds, TrEx achieved 85.51% matched trees while VWF and watershed segmentation had percent matched trees less than 60%. The tree segmentation algorithms had a remarkably consistent accurate tree height estimate with Agisoft point clouds of about 93%. For Pix4D, the highest percent accurate tree height was 91.46% and achieved by the VWF algorithm. Watershed segmentation and TrEx algorithms had 89.98% and 89.02% accurate tree height, respectively, with PiX4D point clouds.