Development of a Machine Learning Method for 3-D Point Cloud– Based Digital Transformation of Forest Resource Survey

Abstract

Laser measurements and aerial photographs have facilitated the modeling of forest structures. A classifier for tree trunks based on PointNet++ was developed for automatic forest surveys using point clouds obtained from ground laser measurements. The study sites were selected from plantation forests in Fukushima Prefecture and secondary forests in Miyagi Prefecture. Machine learning of the classifier was performed under various learning conditions using a manually segmented point cloud as training data. The discrimination rate was obtained from both training and validation datasets. The classifier demonstrated effective accuracy in artificial forests. However, in secondary forests, the discrimination rate and generalization performance were lower because of the complexity of the forest structure compared with that in previous studies. Still, the accuracy was sufficient for estimating tree positions and diameters.