Evaluation of Machine Learning Algorithms for Classification of Infrastructure Elements in Complex Structures

Abstract

3D as-built models play a crucial role in urban dynamics within technical infrastructure projects. Enabling 3D data analysis in dynamic urban areas is advantageous for construction processes and management tasks, including production monitoring and layout planning. Point clouds represent a scene by deploying a collection of 3D points that include both the position and color information of each point. It is crucial to assign semantic information to these points, known as point cloud classification (PCC), for the documentation and monitoring of the 3D environment. Machine learning (ML) classifiers that use multiscale geometric features of each point are frequently utilized for point cloud classification (PCC). This study aims to assess the classification performance of ML classifiers in complex infrastructure domains and determine the geometric characteristics that are most effective in identifying target scenes. Random Forest (RF), eXtreme Gradient Boosting Machines (XGB), and Light Gradient Boosting Machines (LGBM) classifiers were utilized in two case studies. Based on the experimental tests, XGB demonstrated greater accuracy than RF and LGBM. Moreover, despite employing distinct approaches, these classifiers were found to generally agree on the importance of specific geometric features. Our findings offer guidance for classifying technical infrastructure elements in real-world scenarios.