Terrestrial Laser Scanning (TLS) for tree structure studies: a review of methods for wood-leaf classifications from 3D point clouds

In the last decades new remote sensing and proximal sensing techniques have been developed for forest monitoring and vegetation mapping application. In this context LiDAR systems are considered one of the most accurate remote/proximal sensing technologies. Ground-based LiDAR systems (Terrestrial Laser Scanner-TLS) have emerged, in the last twenty years, as valid alternatives to traditional ground-based forest inventory techniques. It represents a non-destructive approach to estimate canopy and stem volume and biomass-related parameters with a degree of detail greater than that achievable by traditional fieldwork. A particular application of the TLS in forestry studies concerns the identification and separation of wood components from non-wood parts of trees because it is a fundamental and essential step to correctly estimate several attributes of trees and crowns. Over the past decade, thanks to significant advances in technology and improved analysis methods, several approaches have been proposed to separate different materials within the same TLS point cloud: methods that rely on geometric features of the TLS data; methods that rely on radiometric features of the scanned points; and methods that exploit both geometric and intensity features. This article analyses the methods developed over the last twenty years to classify woody and leaf components from TLS data. The study highlights advances in TLS technology and the evolution of analytical methods, describes main characteristics of each approach, and give information on their main application in forestry studies. In general, the most recent approaches are based on the geometric characteristics of point clouds. Recently, the use of machine learning techniques has become widespread and has proven effective in separating leaves from wood in TLS data. In recent years, we have observed an increase in the use of classifiers based on artificial neural networks that seem to be able to achieve high-precision separation results.