Boosting Digital Heritage with AI to Derive Enlightening Information

The success of digital technologies in the heritage domain has led to exponential growth in 3D digital heritage data in both volume and resolution, necessitating new solutions capable of processing ever-larger datasets. Advancements in artificial intelligence (AI), particularly in machine learning (ML) and deep learning (DL), provide powerful tools to address this challenge, offering significant advantages in speed and accuracy while assisting domain experts with documentation, preservation, conservation, and interpretation. The application of these advancements to the heritage field, however, is very limited and presents several challenges due to the peculiar characteristics of heritage data, including challenging surfaces for 3D reconstruction, inadequacy of traditional algorithms for archaeological feature detection in LiDAR data, scarcity of annotated training data, and varying or unique class definitions that hinder model generalisability across diverse heritage contexts. This thesis presents original ML/DL-based contributions to address these challenges: • A methodological framework for assessing Neural Radiance Field (NeRF) capabilities in heritage-specific scenarios is presented through comparative evaluation against conventional photogrammetric pipelines, with heritage assets contributed to the NeRFBK benchmark to enable reproducible evaluation. • A Multi-Level Multi-Resolution (MLMR) approach for generating archaeological Digital Feature Models (DFMs) from LiDAR data, subdividing the task into two resolution levels to enable progressive validation and task-specific optimisation, employing a deep learning backbone and a train-test consistency strategy, validated across Mediterranean and Alpine archaeological sites. • Two methodological approaches integrating 2D and 3D data for the semantic enrichment of heritage 3D data, with comparisons against fully 3D-based approaches: the first incorporates DSM-derived depth maps and existing technical drawings to automatically generate training data for image-based DL models and segment planar surfaces by back projecting the results onto the point cloud; the second exploits photogrammetric image redundancy for large-scale training data generation and planar surface segmentation, employing voxelisation and voting mechanisms to assign final labels to each point, with large-scale validation on the Colosseum dataset. • Finally, a zero-shot, multimodal large language model (MLLM)-based approach for segmenting heritage classes from 3D mesh textures without domain-specific training data is validated on Trajan's Column in Rome and the Borobudur temple in Indonesia. Real-world applications demonstrate the concrete impact of the proposed methods, including the discovery of previously unknown archaeological features at Rusellae and Moscona, and the successful upscaling of the proposed 2D/3D framework to the Colosseum dataset, demonstrating its applicability at a large scale. Beyond empirical validation, this research contributes to the growing body of literature bridging the heritage and AI domains, providing methodological frameworks and benchmark contributions that address the specific characteristics of heritage 3D data and support the progressive automation of heritage documentation workflows.