<p>Model performance in the flint/obsidian class.</p>

The quantification and identification of components in archaeological micromorphology remain subjective and challenging, particularly for early-career researchers. To address this, we developed a deep learning tool for the automatic segmentation of three materials commonly found in Palaeolithic contexts and thin sections: bone, charcoal, and lithic fine-grained debitage (flint and obsidian). Using high-resolution photomicrographs of 57 thin sections in plane-polarised and cross-polarised light, we trained and evaluated state-of-the-art convolutional neural networks (CNNs) for material segmentation. The best-performing configuration, a U-Net with an InceptionV4 encoder, achieved mean intersection over union (IoU) scores of 0.96 for flint/obsidian, 0.80 for bone, and 0.82 for charcoal. The models also classified the relative abundance of each material with balanced accuracies of 0.99 for flint/obsidian, 0.92 for bone, and 0.85 for charcoal. These results demonstrate the potential of deep learning to enhance objectivity, accuracy, and reproducibility in archaeological micromorphology, providing a valuable resource for future geoarchaeological research.

考古微形态学中组分的定量与鉴定仍存在主观性强、操作难度大的问题，尤其对于早期职业研究者而言。为解决这一难题，我们开发了一款深度学习工具，可自动分割旧石器时代考古遗存与薄片中常见的三类物质：骨、炭屑以及细粒石质碎屑（燧石与黑曜石）。我们利用57份薄片在单偏光与正交偏光照明下拍摄的高分辨率显微照片，训练并评估了当前最优的卷积神经网络（Convolutional Neural Networks, CNNs）以完成物质分割任务。表现最优的模型架构为搭载InceptionV4编码器的U-Net，其在燧石/黑曜石类别上的平均交并比（Intersection over Union, IoU）得分为0.96，骨类别为0.80，炭屑类别为0.82。该模型还可对各物质的相对丰度进行分类，其平衡准确率在燧石/黑曜石上为0.99，骨为0.92，炭屑为0.85。上述结果证明了深度学习在提升考古微形态学研究客观性、准确性与可重复性方面的潜力，可为未来地质考古学研究提供宝贵的资源。