Machine learning for cultural heritage conservation: decoding the past through analysis of hyperspectral data

Cultural heritage conservation and restoration stand at the crossroads of art, history, science, and technological progress. Traditional conservation and restoration methods, while invaluable, are often constrained by their ability to adapt to the complexity of cultural artefacts. In the contemporary landscape, analytical instruments and computational technology are profoundly shaping the interdisciplinary field. Machine learning and data analysis, with their capacity to unravel intricate patterns and trends within extensive datasets, offer a promising avenue for enhancing conservation and restoration practices. This thesis aims to investigate the potential and limitations of machine-learning techniques in processing extensive hyperspectral data acquired from historical art objects. Through various case studies presented, we assess the effectiveness of machine learning models, from off-the-shelf techniques to self-developed algorithms, in supporting tasks ranging from material diagnostics and classification to mapping and digital restoration. Additionally, we critically evaluate the challenges and limitations associated with the implementation of machine learning, specifically constrained by CH, and explore the transferability of machine learning models to similar scenarios. Rooted in the obtained results, this thesis contributes to the ongoing dialogue on leveraging cutting-edge technologies to preserve and celebrate our diverse cultural heritage.