Deep learning-based identification of carbonized seeds: A case study on Panicum miliaceum (Broomcorn Millet) and Setaria italica (Foxtail Millet)

Abstract

Carbonized seeds recovered from archaeological sites are commonly identified through manual inspection, a time-consuming and low-throughput process. In particular, many fragmented carbonized seeds cannot be accurately identified by the naked eye. Among these, carbonized millet seeds are especially challenging due to their morphological similarity when charred. This study is the first in archaeobotany to apply and evaluate deep learning techniques for the automatic identification and species-level classification of carbonized millet seeds. A dataset comprising 3,000 images was constructed using modern Panicum miliaceum (broomcorn millet) and Setaria italica (foxtail millet) seeds in three different seed-condition image datasets: intact, fragmented, and carbonized. Three deep learning models—ResNet50d, YOLOv5s, and ViT-Tiny—were assessed systematically. The ResNet50d model performed best, and did so using carbonized seeds as a tranining set, achieving 95 % accuracy for distinguishing carbonized broomcorn millet and foxtail millet after a mere 3-hour training period, demonstrating its suitability for identifying the carbonized millet seeds. Notably, training with only 300 images per carbonized seed type (600 images in total) yielded stable F1-scores exceeding 90%, significantly reducing training costs. Comparative analyses across different seed-condition image datasets and image qualities revealed that carbonized seeds provided the most discriminating features. Removing image backgrounds further improved model accuracy. Grad-CAM visualizations confirmed both the effectiveness and interpretability of the ResNet50d model. This study highlights the transformative potential of deep learning for automating carbonized seed identification at the species level, offering a practical and efficient approach for the analysis of archaeobotanical remains.