This study explores the use of convolutional neural networks (CNNs) to classify archaeological obsidian artifacts by their geological source using standard documentation photographs. Drawing on a dataset of images taken under varied field and laboratory conditions, we train and evaluate multiple CNN architectures to assess the feasibility of this approach as a low-cost alternative to geochemical sourcing. The models achieve high precision and recall for several well-represented sources, and Grad-CAM visualizations indicate that classification is often based on visually meaningful surface features. This work represents a first step towards developing a new sourcing methodology, as the technique demonstrates strong potential for scaling up obsidian sourcing in contexts where access to laboratory equipment is limited or cost-prohibitive. We argue that with a larger and more diverse image dataset—including a broader range of artifact types and source locations—image-based classification could become a practical and accessible tool for archaeological research in Middle America and beyond. • Convolutional neural networks used to source images of obsidian artifacts. • Standard documentation photos of obsidian tested with high classification accuracy. • Promising results imply wider applicability to additional sources and regions. • First step towards low-cost method democratizing sourcing of obsidian assemblages. • Only standard camera and internet connection required to source obsidian.
Lyons et al. (Sat,) studied this question.