Optimising 3D point cloud semantic segmentation: ML and manual refinement in the unesco saltaire village

Semantic segmentation of 3D point clouds has been an ongoing challenge in recent years. The present research focuses on refining existing standalone Machine Learning (ML) algorithms to enhance their performance in segmenting mid-19th-century industrial housing architectural components, drawing on the UNESCO World Heritage Site Saltaire Industrial Village. Its architecture is actively influenced by contemporary human activity, introducing complexities into the original fabric and spatial composition. This research provides methodological insights into optimising segmentation performance through a combination of pragmatically reviewed ML classification techniques and manual refinement strategies. The methodology is based on two classification methods: Standalone ML and Multilayer ML. For the first time, this study provides detailed evidence of the challenges encountered in transitioning from traditional human-led models to HBIM in densely altered heritage environments. Results evaluate the performance of each method with the final aim of laying the groundwork for a semi-automated AI-backed scan-to-BIM approach for 19th-century architecture, contributing to a deeper understanding of its unique characteristics and supporting a sustainable transition to robust HBIM.