Towards interpretable deep learning in ceramic petrographic fabric classification through a comparative study of convolutional neural networks and vision transformers

In recent decades, there has been growing interest in the application of Artificial Intelligence (AI) to archaeology and archaeometry, where it shows promising potential in addressing specific research challenges and supporting complex analytical tasks. Among these, an innovative development is represented by the application of Machine Learning (ML) and Deep Learning (DL) models for ceramic analysis. Recent studies have indeed demonstrated their effectiveness in identifying specific compositional, technological, or stylistic patterns (Qi et al., 2022; Ruschioni et al., 2023). This study explores the use of Deep Learning models for the classification of Levantine ceramic thin sections by identifying distinctive compositional features. A dedicated dataset of ceramic thin section images was built to train and validate the models. The ceramic samples are dated back from the Early Bronze Age to Iron Age and 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). This research applies both Convolutional Neural Network (CNNs) and Vision Transformer (ViTs) models to evaluate their performance in classifying ceramics within their petrographic fabrics, which reflect the raw materials used and the production technologies. Additionally, Explainable Artificial Intelligence (XAI) techniques are employed, with the aim of enhancing transparency and interpretability (Zhong et al., 2022). Furthermore, this study investigates the implementation of Adaptive Vision Transformer (A-ViT) and Dynamic Vision Transformer (DynamicViT) architectures, which introduce adaptive token selection mechanisms (Liu et al., 2024). These methods dynamically allocate computational resources by focusing attention on the most informative regions of an image, allowing the models to improve both classification accuracy and efficiency. The use of these advanced transformer-based models, in conjunction with XAI, aims to open new pathways for the application of interpretable deep learning in archaeological science.