Explainable vision transformers for the petrographic classification of Levantine ceramics

The archaeological context of the Levant is particularly rich and complex, as numerous sites emerged starting from the Early Bronze Age, playing a strategic role in trade, cultural exchange, and technological innovation. The study and classification of ancient ceramics is fundamental to understanding the social and cultural identity of this region, offering important insights into technological practices, provenance of raw materials, and interactions between ancient societies. This study investigates the use of Deep Learning (DL), specifically Vision Transformers (ViTs), for the automated classification of Levantine ceramics based on their petrographic fabrics. A dataset composed of ceramic thin section images dated back to the Early Bronze Age and Iron Age was built to train and validate the models. The ceramic samples come from several archaeological sites across the Levant, including Bethlehem (West Bank), Tell el-Far’ah North (West Bank), Khirbat Iskandar (Jordan), Khirbat al-Batrawy (Jordan), Ebla (Syria), Jericho (West Bank), Tell Nebi Mend (Lebanon) and Tell Qasile (Israel). In particular, the study explores the application of AviT and DynamicViT models, which implement adaptive token selection to reduce computational cost while focusing on the most relevant image regions. These models are evaluated not only in terms of classification accuracy but also for their interpretability. Moreover, to improve model transparency, the study applies several explainability (XAI) techniques. In addition to visual tools such as attention maps, it also incorporates advanced semantic-probabilistic methods which allow a more structured and interpretable understanding of model behavior. By combining high-performing ViT models with diverse XAI strategies, this project aims to develop a robust and transparent framework for the classification of ancient ceramics, bridging artificial intelligence with archaeometric studies.