Lidar-based Species Classification Using Multivariate Cluster Analysis

Understanding that various tree species have characteristics similar to each other, it follows that some type of hierarchical classification scheme could be used to identify species using LIDAR data. Cluster analysis, one of the unsupervised classification methods, was conducted for all individual trees using the k-medoid algorithm. Instead of using one-step cluster analysis, a stepwise cluster analysis was developed based on the statistical criteria to test hierarchical relationships between species. Two seasonal LIDAR datasets collected at the Washington Park Arboretum in Seattle, Washington were used for this study. Parameters derived from structure and intensity measurements using two LIDAR datasets were used for the stepwise clustering analysis. This paper shows that a variety of tree species can be naturally clustered with a hierarchy using LIDAR-derived structure and intensity measurements. Stepwise cluster analysis showed that the species with similar characteristics seem to be clustered into a single group while the species with different characteristics are likely to be clustered into different groups based on the reliable statistical criteria. The clustering results using different seasonal datasets revealed that using both seasonal datasets clustered species more reasonably than using either one of the datasets. When using only leafon data, the structure of clusters was not reasonably formed even at the first step of cluster analysis. It should be noted that the clustering results would vary depending not only on the variables used but also on the selected species groups or the number of individual trees.