Deep learning enables automated detection of dinosaur footprints with high accuracy

Abstract Dinosaur footprints provide crucial paleoecological information about locomotion, behavior, and tracksite distributions. Traditional tracksite surveys rely on time-consuming manual identification by expert researchers, with results influenced by preservation quality and subjective interpretation. Here, we present an automated detection system for dinosaur footprints using You Only Look Once version 8 (YOLOv8), an effective object detection neural network. We trained the model on 49,242 images from the AI-Hub dataset, comprising theropod, ornithopod, and sauropod footprints from Korean tracksites. The final model achieved a mean average precision (mAP50) of 0.949 and mAP50-95 of 0.660. To address resolution-dependent detection challenges, we developed a multi-scale detection approach combining predictions across eight different image resolutions. Testing on previously unseen tracksites demonstrated successful detection of footprints from both Korea and Brazil, including the ichnospecies Farlowichnus rapidus , indicating strong cross-regional generalization. Our results show that detection accuracy depends on illumination conditions, preservation quality, and the presence of outline markings. This deep learning approach offers an efficient alternative to manual tracksite exploration, enabling rapid spatial distribution mapping of extensive dinosaur tracksites while providing objective, reproducible detection results.