Tree species classification using intensity patterns from individual tree point clouds

Personal laser scanning has evolved into a cutting-edge technology for obtaining fast and accurate biometric measurements of individual trees in a forest. However, recent studies assessing tree species labels on single tree point clouds have been insufficiently accurate in complex forest ecosystems; moreover, explainability of machine-learning methods used in published studies has been insufficient. Whether the predictions of black-box models suffer from over-fitting or whether they are based on characteristic species traits often remains unclear. To solve this problem, we present a simple classifier combining random forest models with decision rules, trained on 9 common tree species in Central Europe. Explainable elements are a soft classifier on classification probabilities and detailed analysis of variable importance and minimal variable depth. The overall classification accuracy was 89.8% for nine species, with greater values for the four major species (spruce, pine, oak, and beech). Intensity measures in the upper tree section and tree geometry ratios were the most important predictors. The method proposed in this study can potentially be used to analyze forest ecosystems in more spatial detail by addressing species-specific research questions to an unprecedented degree.