Tree genera classification using airborne LiDAR data by ensemble methods

We propose an ensemble classification method for classifying tree genus by using LiDAR (Light Detection and Ranging) data. We have developed a set of descriptors (features) related to the geometric information given by the point cloud. The second set of features is derived from a more conventional method and is related to the vertical point distribution of the point cloud. We built two classifiers separately (geometric classifier and vertical profile classifier) using the two sets of features and then combine the classifiers for improving overall classification accuracy. Our study area is located north of Thessalon, Ontario, Canada and we are classify trees into three genera (pine, poplar and maple) within our study sites. Result show that the average classification accuracy for the geometric classifier is 88.0% and 88.8% for vertical profile classifier. When the classifiers are combined, the overall accuracy improved to 91.2%. Background and Relevance The use of aerial LiDAR in forestry applications has become increasingly popular for its ability to acquire 3D information and has proven successful in tree species/genera classification (Holmgren and Persson 2004; Brandtberg 2007; Holmgren et al., 2008; Kato et al., 2009; Ørka et al., 2009; Vauhkonen et al., 2009; Korpela et al., 2010 and Kim et al., 2011). Our first set of features are derived from the geometry of the LiDAR point distribution, this approach can be found in Kato et al. (2009) where the authors fit curved surfaces to the individual LiDAR tree crown and Vauhkonen et al. (2009) compute alpha shapes of the LiDAR tree crowns. Both methods derive features related to the outer shape of the tree crown, we further develop features that relate to the outer as well as inner geometry of the tree (branching levels). The second set of features is calculated from a more convention approach, examples of such an approach include Holmgren and Persson (2004); Brandtberg (2007); Ørka et al. (2009); Korpela et al. (2010) and Kim et al. (2011). These authors derived features from the vertical point profile reflected from the tree (or tree crown) and calculate statistical metrics that summarizes the point distribution within specific height percentiles or the entire profile. The advantage of the geometric features is that they can be easily related to the physical and biological implication of tree form, however they are usually more computationally expensive. Conversely, vertical profile features are computationally efficient but are less intuitive. This research takes advantage of the both perspectives and combines both classifiers to yield a better result.