Deep Learning for ancient ceramic classification: saliency maps as a tool for models interpretability

Ancient ceramic artifacts are important remains within archaeological contexts. They represent indicators of social and human activity, giving information on production techniques, technological development, and raw materials provenance to reconstruct relationships and trade routes among past societies [1]. Tracing ancient trade interactions is still challenging due to the subjectivity and the time-consuming nature of the current procedure, which is mainly based on minero-petrographic analyses of pottery. In the last decade, there has been growing interest in the application of automatic methodologies for ceramic grouping. The use of classification algorithms has significantly contributed to recognize specific compositional, technological or stylistic patterns [2, 3]. The present study aims to classify Levantine ceramic thin sections using Deep Learning (DL) models, specifically Convolutional Neural Networks (CNNs) and Vision Transformers (ViT). The objective was to group these ceramic samples into their respective petrographic fabrics and compare the results of both models to evaluate their classification effectiveness. A further goal was to provide a partial interpretation of the CNN and ViT model results by applying model processing based on saliency maps (heat maps) to visualize the distinctive features that contributed to the classification of the samples [4]. The results are promising and highlight the importance of the application of saliency maps, which help identify the specific elements of the input data that are most relevant in determining the output.