BUILDING A HARDWOOD/SOFTWOOD TREE CLASSIFIER USING LIDAR AND GUELPH’S URBAN TREES

Since it’s introduction to forestry in the ‘80s, the use of light detection and ranging (LiDAR) has expanded to encompass many functions desirable to forest managers. One function yet to be fully optimized is the detection and classification of individual trees in complex forest stands, which stands to be a useful tool for the management of forests by acting as a proxy for timber scaling, mapping sensitive habitat, or spatial modeling. The main objective of this study was to create a workflow in ESRI ArcPro GIS software to process classified LiDAR point cloud coordinate data in order to isolate LiDAR returns associated with individual trees and output a data table of metrics derived from the LiDAR point cloud corresponding to those trees. A secondary objective of the study was to evaluate which of these LiDAR derived metrics were best suited to differentiate between hardwood (HWD) and softwood (SWD) trees for species classification purposes. A sample of ~230 trees was delineated from publicly available LiDAR data, and tree type was validated by referencing the City of Guelph’s tree inventory data. Three-dimensional tree “hulls” were extracted from the LiDAR data using an existing open-source tree segmentation algorithm executed in the R programming language. A series of descriptive statistics of the distribution of LiDAR returns associated with individual hulls were subsequently used as input variables in a Random Forest (RF) classifier to classify hardwood vs. softwood trees. Classification error rates for “out of bag samples” were 0.1% and 9.5% for hardwoods and for softwoods, respectively. A series of principal component analyses (PCA) were subsequently undertaken using our RF training dataset in order to better understand which were the most important variables to distinguish hardwoods and softwoods. PCAs were first used to select variables from families of like variables to include in a final PCA that included the following variables: mean height (zmean), 75th percentile height (zq75), cumulative percentage of height returns in the 5th layer (of 10) (zpcum5), pecent 1st returns (p1th), cumulative percentage of intensity returns below the 70th percentile height (ipcmz70), and the total area of the tree hull in m2 (Shape_Area). The variance explained by the first and second principal components were 45.22%, and 18.43%, respectively; with tree type best distinguished along PC2. Eigenvectors for the area of tree hulls and intensity of LiDAR returns were most strongly associated with PC2. A visual inspection of aerial imagery corresponding to tree hulls showed that individual hardwoods were often segmented into multiple hulls, whereas softwoods rarely were, which may explain why softwoods were positively associated with the area of tree hulls. We hypothesized, based on relatively high commission errors for hardwood tree hull segmentation, that the LiDAR data were acquired during leaf-off conditions. If true, a lack of foliage to intercept returns and a relatively high proportion of ground returns may explain why LiDAR return intensity was more positively associated with hardwood hulls. We recommend that LiDAR data for species classification be acquired during leaf-on conditions to minimize errors of commission during hardwood tree segmentation, and better distinguish tree species based upon the LiDAR point cloud distribution associated with tree hulls.