Machine Learning-Based Detection of Stone Degradation using TLS, Photographs and HBIM: A Case Study on the Lausanne Cathedral


Loading...

Date

2025

Publication Type

Conference Paper

ETH Bibliography

yes

Citations

Altmetric

Data

Abstract

Historic sandstone structures are vulnerable to environmental degradation, requiring efficient damage assessment and documentation for effective conservation. Traditional manual mapping methods are time-intensive, cost-intensive, and challenging to standardize. Herein we present a semi-automated workflow for damage classification and integration into a historic building information model (HBIM) using terrestrial laser scanning (TLS), orthophotos, and machine learning. A Random Forest model, optimised through forward feature selection and Bayesian hyperparameter tuning, classifies degradation types (contour scaling, biodeterioration, black crust formation and exfoliation) based on geometric and radiometric features. Finally, degradation information is linked to individual stone blocks in the HBIM, and degradation maps can be generated. In a case study on the Lausanne Cathedral, we achieve an overall classification accuracy of 80% using this approach. The results highlight the potential of using machine-learning techniques with TLS data for an efficient and scalable heritage condition assessment.

Publication status

published

External links

Editor

Book title

STONE 2025 - Proceedings of the 15th International Congress on the Deterioration and Conservation of Stone, Volume 1 & 2

Journal / series

Volume

Pages / Article No.

565 - 571

Publisher

15th International Congress on the Deterioration and Conservation of Stone

Event

15th International Congress on the Deterioration and Conservation of Stone (STONE 2025)

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

HBIM; LiDAR; point clouds; degradation mapping; machine learning; cultural heritage

Organisational unit

03964 - Wieser, Andreas / Wieser, Andreas check_circle
03891 - Flatt, Robert J. / Flatt, Robert J. check_circle

Notes

Conference lecture held on September 10, 2025

Funding

Related publications and datasets