Key Variables Affecting the Success of Machine Learning Algorithms in Individual Tree Species Classification in Mixed Forest Ecosystems Using LiDAR Data Collected by Autonomous Drones

Accurate classification of individual tree species in mixed forest ecosystems is a fundamental requirement for forest inventory, biodiversity monitoring, carbon stock estimation, and silvicultural planning. In recent years, LiDAR sensors integrated into autonomous drone platforms have offered new opportunities in this field, providing higher point density, lower cost, and flexible flight planning advantages compared to conventional airborne or terrestrial laser scanning systems. However, the species classification accuracy of machine learning algorithms applied to drone LiDAR data varies significantly depending on numerous variables, including point cloud density, flight altitude, seasonal timing, individual tree segmentation quality, feature selection, algorithm type, and the structural complexity of the forest. This study aims to systematically determine the relative effects of these variables on classification accuracy and to establish a ranking of their importance. The research was conducted with a multi-altitude (40–120 m) and multi-seasonal (leaf-on/leaf-off periods) drone LiDAR data acquisition design in study sites covering at least five coniferous and broadleaved species in temperate zone mixed forest areas. Structural, geometric, and density-based features extracted from point clouds were comparatively tested using Random Forest, Gradient Boosting, Support Vector Machines, k-Nearest Neighbor, and PointNet++ algorithms. Permutation importance and SHAP analyses revealed that seasonal timing and point cloud density were the variables most strongly affecting classification success. The combined use of leaf-on and leaf-off period data significantly increased overall accuracy compared to single-season data. Species-level analyses demonstrated that broadleaved species were classified with higher accuracy compared to conifers, while the highest confusion rates were observed between coniferous species pairs with similar crown structures. The findings provide concrete recommendations for data acquisition strategy, algorithm selection, and feature engineering for operational forest inventory applications.