Spatial Analysis of Dense LiDAR Point Clouds for Tree Species Group Classification Using Individual Tree Metrics

Abstract

This study presents a method of tree species classification using individual tree metrics derived from a three-dimensional point cloud from unmanned aerial vehicle laser scanning (ULS). In this novel approach, we evaluated the metrics of 1045 trees using generalized linear model (GLM) and random forest (RF) techniques to automatically assign individual trees into either a coniferous or broadleaf group. We evaluated several statistical descriptors, including a novel approach using the Clark–Evans spatial aggregation index (CE), which indicates the level of clustering in point clouds. A comparison of classifiers that included and excluded the CE indicator values demonstrated their importance for improved classification of the individual tree point clouds. The overall accuracy when including the CE index was 94.8% using a GLM approach and 95.1% using an RF approach. With the RF approach, the inclusion of CE yielded a significant improvement in overall classification accuracy, and for the GLM approach, the CE index was always selected as a significant variable for correct tree class prediction. Compared to other studies, the above-mentioned accuracies prove the benefits of CE for tree species classification, as do the worse results of excluding the CE, where the derived GLM achieved an accuracy of 92.6% and RF an accuracy of 93.8%.