Machine Learning for UAV and Ground-Captured Imagery: Toward Standard Practices

Our collaborative work began in 2019 with the intent to overcome obstacles that had arisen from the inability to access curated artifact collections from remote locations. It was our specific aim to not only create digital twins of excavated objects that could not be exported out of their country of origin, but also to emphasize the contextual associations of objects residing in hidden museum collections using a range of digital techniques. As part of a growing field project in 2022, machine learning (ML) with YOLOv5, a family of compound-scaled object detection models trained on the COCO dataset was used to classify visual data and advance our understanding of in situ archaeological phenomena prior to destructive fieldwork. While not the sole contribution, the use of object-based machine learning improved quality and range of information obtained in non-destructive site surveys and improved data sharing capacity. Despite challenges encountered while training the algorithm and classifying objects, combining ML with drone data collection will continue as part of our long-term spatial data recording procedure. Despite both success and failures reported here, this work contributes to greater standardization of ML techniques in archaeological practice.