EXPLAINABLE DEEP LEARNING FOR PETROGRAPHIC FABRIC CLASSIFICATION OF LEVANTINE POTTERY

In recent decades, Artificial Intelligence (AI) has become a powerful tool in the field of archaeology. In particular, Machine Learning (ML) and Deep Learning (DL) techniques have been applied to the study of ceramics. Several algorithms have been tested for pottery classification, based on typology, decoration, and shape investigation or through geochemical analysis for provenance studies [1,2]. Automated methodologies represent an innovative approach for resolving classification tasks and appear to contribute to the recognition of specific compositional, technological, or stylistic patterns [3,4]. This study focuses on the petrographic fabric classification of Levantine ceramics using two DL approaches: Convolutional Neural Networks (CNNs) with a ResNet18 architecture and Vision Transformers (ViTs). The aim is to investigate the potential of these techniques in ceramic petrography, specifically evaluating their effectiveness in classification tasks. To train and validate the models, ceramic samples dated back to the Early Bronze Age and Iron Age from Bethlehem (West Bank), Tell el-Far’ah (West Bank), Khirbat Iskandar (Jordan), Khirbat al-Batrawy (Jordan), Ebla (Syria) and Jericho (West Bank) were selected, creating a dataset representative of Levantine pottery production. Furthermore, this research integrates Explainable AI (XAI) using visual explanation tools, such as saliency and attention maps, to reach a deeper understanding of the classification process by visualizing the most important features used by CNNs and ViTs. ViTs have never been applied to petrographic fabric classification, and XAI has not yet been utilized to clarify model predictions in this domain. In this context, results are significant, demonstrating the potential of these automated methodologies. The application of explainable approaches made the models more transparent and evidenced the importance of selecting representative images, where specific minerals and patterns of each petrographic fabric can be observed. The saliency and attention maps highlighted the inclusions that the models considered relevant for classification, helping in refining the selection of training images and leading to more robust classification models.