Vegetation filtering in photogrammetric 3D point cloud data by tree-based machine learning models: a comparative study

Abstract

Identifying and filtering vegetation from photogrammetricbased point cloud data are required for many applications, such as environmental monitoring, urban planning, forestry and hazard management. This work presents a comprehensive
study on point classification using advanced tree-based machine learning models, including Decision Tree Classifier, Random Forest, AdaBoost, XGBoost, and CatBoost. Six different datasets are utilized for comparison between the machine learning models. Random Forest classifier with hyperparameter tuning outperforms other models, demonstrating superior precision in filtering vegetation points. Visible-band vegetation indices with fixed thresholds are also evaluated, but their accuracy is lower than that of machine learning models, despite being easier to implement. The findings not only show the performances of tree-based machine learning models for filtering vegetation in photogrammetric 3D point cloud data, but they also suggest promising potential to transform the binary classification task into a multi-class classification to achieve higher granularity.