DIGITAL.CSIC. Repositorio Institucional del CSIC
Publicación
Artículo científico (JournalArticle)
Evaluation of Automated Machine Learning and Model Transferability Approaches for Automating the Diagnosis of Historic Constructions. A Case Study in the Almeida Fortress (Portugal)
23 páginas, 23 figuras, 14 tablas, This paper assesses the potential of Automated Machine Learning (AutoML) and pretrained
algorithms for automating information extraction from 3D point clouds in heritage conservation.
The study evaluates the feasibility of identifying essential structural elements and recurrent
damage patterns in buildings with similar geometry and deterioration characteristics. AutoML is
applied to optimize algorithm selection and hyperparameter tuning for constructive segmentation
of 3D point clouds, outperforming traditional Random Forest classifiers. Additionally, pretrained
algorithms such as Fuzzy K-means, CANUPO, and RANSAC Shape Detection efficiently identify
damages like deformation, material loss, moisture areas, salt deposits, and biological growth,
requiring minimal training data. This methodology was validated on two iconic structures of the
Almeida Fortress in Portugal: the Master Gates of San Francisco and San Antonio. Training with
approximately 25% of the San Francisco Gate dataset enabled accurate prediction of both
structural elements and damage patterns across the remaining datasets., The work was supported by the Comunidad de Madrid
[APOYO-JOVENES-21-RCDT1L-85- SL9E1R]; Ministerio
de Ciencia, Innovación y Universidades [PID2024-
159276OA-C44]; Ministerio de Ciencia, Innovación y
Universidades [PID2022-140071OB-C21], Peer reviewed