Automated Detection of Protohistoric Tumuli in the Friulan Plain: A LiDAR-Based Deep Learning Approach with Mask R-CNN and U-Net

This thesis investigates the application of deep learning frameworks to automate the detection of protohistoric tumuli using LiDAR-derived digital elevation models in the Upper Plain of Friuli-Venezia Giulia. By leveraging the Visualization for Archaeological Topography (VAT) method, high-resolution Li-DAR data are transformed into composite imagery that serves as the basis for model training and analysis. Two state-of-the-art pre-trained convolutional neural networks—Mask R-CNN for instance segmentation and U-Net for semantic segmentation—are adapted via transfer learning to identify and delineate burialmounds. The study evaluates these models on annotated datasets, highlighting their respective strengths and weaknesses in detecting archaeological features. Comparative analysis shows that while Mask R-CNN effectively segments closely spaced anomalies, U-Net adopts a more conservative approach with fewer false positives, thereby providing a more accurate delineation of tumuli shapes. Both models are subsequently deployed on unannotated imagery to identify new potential sites, demonstrating the practical utility of deep learning frameworks in guiding future field surveys. Overall, these findings underscore the promise of integrating deep learning with remote sensing techniques for archaeological prospection and emphasize the importance of expert interpretation in refining automated detections. These results lay the foundation for further research into optimized training strategies and model architectures tailored to the challenges of archaeological feature detection.