Automatic damage detection in 19th–20th century heritage buildings using R-C-C fusion machine learning with 3D laser scanning

This study presents an automated structural damage detection methodology for 19th–20th century heritage buildings using a Roughness–CANUPO(1)–CANUPO(2) (R–C–C) machine learning algorithm combined with 3D laser scanning. To address the limitations of traditional inspection methods in heritage conservation, a non-destructive testing (NDT) integrating surface roughness analysis and machine learning was applied to six heritage buildings constructed with red brick, limestone, and terracotta. High-resolution point cloud data(PCD) were acquired using terrestrial laser scanning, and Local Neighbour Radius (LNR) values were optimised to maximize the separation of crack and wall surface features during roughness-based filtering. A two-stage CANUPO classifier based on the support vector machine learning (SVM), trained on roughness-derived features, was employed to automatically distinguish cracks from undamaged surfaces and joints. Experimental results confirmed that optimal LNR and filtration ratio tuning were essential for effective crack visibility and classification performance. Specifically, under optimised conditions, maximum crack visibility reached 47.28% and 32.74% for red brick walls, 63.48% and 30.23% for limestone walls, and 82.56% and 30.34% for terracotta columns. These results highlight the importance of adapting LNR values and filtering strategies to material-specific surface geometries, particularly in curved components like terracotta columns where 3D curvature influences roughness behaviour. The R–C–C approach enables scalable and accurate structural condition assessment without physical contact, offering a practical tool for the structural monitoring and long-term preservation of historically significant architecture. • Automated structural damage detection methodology for 19th–20th century heritage buildings • Roughness–CANUPO(1)–CANUPO(2) (R–C–C) machine learning algorithm combined with 3D laser scanning • Local Neighbour Radius (LNR) values were optimised to maximize the separation of crack and wall surface features • A two-stage CANUPO classifier based on the support vector machine learning (SVM)