Performance of tree stem isolation algorithms for terrestrial laser scanning point clouds

LiDAR sensors present increasing popularity in the Forestry sector, due to its capability of acquiring high resolution three dimensional data of the forest, useful to a variety of applications. Terrestrial Laser Scanning (TLS) consists of a way to collect large amounts of forest data at the plot level, making further 3D modelling and tree reconstruction possible, thus allowing foresters to extract dendrometric variables with high accuracy from those point clouds. In order to enjoy the full potential of TLS technology for forest inventory, tool sets to extract useful information from TLS point clouds of a variety of tree species are required, starting by stem isolation, which is the core of silviculture and the most targeted outcome of forest management everywhere. The present study aimed to assess the performance of three different methods of stem isolation from TLS point clouds of single trees, both boreal and tropical species. At the same time making the algorithms available as an open source R package. The methods were adapted from three main authors. They rely on finding one main trunk in the point cloud, followed by a circle or cylinder fitting procedure on trunk sections to precisely extract only the stem points. The circle-fit based method had better performance in most cases. Accuracy was higher for all algorithms when tested on boreal trees point clouds, with stem diameter RMSEs ranging from 1.53 to 3.15 cm. For the tropical species the RMSEs ranged from 3.50 to 7.54 cm. Best diameter estimations were obtained for Pinus sylvestris, followed by Picea abies, Eucalyptus sp. and Pinus taeda, respectively. All point clouds had reduced density, keeping less than 300 thousand points per tree, and processing time varied from a few seconds up to 20 min/tree, depending on the method applied and point cloud size.