Comparative evaluation of CNN and ViT architectures for the petrographic classification of Levantine ceramic micrographs

Abstract

This study explores the application of artificial intelligence (AI) techniques for the classification of Levantine ceramic thin sections based on their petrographic fabrics. The Levant is a key archaeological region, particularly during the transition from the Late Chalcolithic to the Early Bronze Age, marked by urbanization, craft specialization, and expanding interregional exchange networks. Ceramic production therefore provides crucial evidence for understanding technological practices and cultural interactions. We apply two deep learning architectures, convolutional neural networks (CNNs) and vision transformers (ViTs), to a large and heterogeneous dataset of ceramic thin-section micrographs dating from the Uruk period through the Bronze and Iron Ages. The dataset includes samples from 14 archaeological sites across the Levant, representing 20 different petrographic fabrics. To evaluate model robustness and generalization, we performed additional experiments by excluding samples from selected sites using them as an unseen test domain, simulating a real-world classification scenario under domain shift and class imbalance. Model performance was evaluated using macro-, micro-, and weighted-F1 scores. To improve model transparency, visual explainable AI approaches, such as Guided Grad-CAM and transformer attention maps, are applied to identify the most informative features driving classification decisions. The results demonstrate high classification performance, reaching an accuracy of 88.44 % for ResNet18 and 90.52 % for DeiT-small. Visual explanations further indicate that the models rely on mineralogical and textural features consistent with petrographic criteria used by human experts, supporting the archaeological interpretability of the proposed approach.