Machine Learning‐Based Automatic Tree Species Detection Using LiDAR Data: Parameter Optimization and Feature Selection

ABSTRACT Today, point cloud data obtained with Light Detection and Ranging (LiDAR) systems enable economical and rapid classification of urban trees using machine learning methods. The determination of optimum parameters and selection of classification features contribute significantly to the success of automatic tree species detection. In this study, we present a method for estimation of optimum parameters and selection of classification features for the automatic detection of deciduous and coniferous tree species in urban areas by machine learning algorithms using LiDAR point cloud. Individual tree crowns were obtained by machine learning‐based mean shift clustering of high vegetation class which was generated through hierarchical rule‐based classification in the urban study area located in Fatih, Istanbul. Reference data for obtained individual tree crowns, including tree species information, were collected by detailed fieldwork. The grid search approach was used to optimize the parameters of Random Forest (RF) algorithm which was used for the classification of obtained tree crowns. After the parameter optimization step, the classification features were evaluated using the Mean Decrease in Gini (MDG) to determine the appropriate features. Tree species were automatically detected as deciduous or coniferous within the urban study area. The RF algorithm achieved 90% overall accuracy with optimized parameters and eight selected features. All machine learning‐based automatic tree species detection processes were carried out using custom‐developed Python code (Python 3.6.4).