Validating predictions of burial mounds with field data: the promise and reality of machine learning

A pre-trained machine learning model, despite achieving high self-reported success rates, failed to accurately detect burial mounds in satellite imagery when validated against field data. The study found false negative and positive rates of 95–96% and 87–95%, respectively, with true positives accounting for only 5–13%. This challenges the optimistic narrative surrounding machine learning (ML) applications in archaeological prospection, highlighting significant limitations in detecting varied features within heterogeneous landscapes. The research underscores the substantial time, expertise, and resources required to improve model performance, questioning the efficiency of ML over manual inspection methods.

Sobotkova, A., Kristensen-McLachlan, R. D., Mallon, O., & Ross, S. A. (2024). Validating predictions of burial mounds with field data: the promise and reality of machine learning. Journal of Documentation, 80(5), 1167–1189. https://doi.org/10.1108/jd-05-2022-0096