An Experimental Analysis of Machine Learning Algorithms and Necessity of Classification System for Semantic Segmentation of Point Clouds for Heritage Buildings

The accurate segmentation and classification of heritage point cloud data is a critical step in developing effective Heritage Building Information Modelling (HBIM) workflows. Thus, this article aims to experiment, evaluate and compare various machine learning (ML) algorithms in point cloud segmentation for heritage conservation. A set of exploratory ML experiments using RANSAC, DBSCAN, HDBSCAN and K-Nearest Neighbors (KNN) is conducted to test initial segmentation performance on a heritage case study. With the key learning from the ML experiments, the research investigates the foundational role of classification systems and their interoperability in enabling ML-based segmentation. Through a structured experimental process, the article evaluates four major classification systems, Uniclass, IFC, ETIM and CCI, for their applicability in heritage contexts and their ability to support semantic labelling within HBIM environments. The research further tests the interoperability of these systems within Autodesk Revit, assessing how well they facilitate data exchange and minimise manual translation during model development. The findings reveal significant variations in segmentation accuracy, computational efficiency and robustness among the evaluated algorithms. It provides insights and key learnings about point cloud segmentation algorithms commonly employed in computational geometry and computer vision, with a focus on HBIM. It also highlights the superiority of specific models in handling complex architectural details typical of heritage structures. While the ML algorithms offered insights into the clustering potential of heritage point clouds, their limitations reinforced the necessity of structured classification schemas. The findings underscore that effective HBIM segmentation begins with well-defined, interoperable classification systems, with ML models serving as tools that benefit from such structured inputs. The novelty of this research lies in its focused application of ML-driven segmentation to heritage conservation, providing valuable insights in digital heritage documentation. This contribution advances the field by offering a benchmark for ML applications in cultural heritage, guiding future developments in automated heritage data processing. The results suggest that while these algorithms can facilitate early-stage segmentation, they lack intrinsic semantic awareness, necessitating further exploration of hybrid ML approaches while the Uniclass classification system is found to be the most suitable taxonomy for semantic classification.