Using deep learning to detect rare archaeological features: A case from coastal South Carolina, USA

Among archaeologists using remote sensing there is tremendous potential for the use of deep learning models for the prospection of archaeological features. The need for relatively large training datasets, technical expertise, and computational requirements, however, has slowed the adoption of these techniques. Here, we train a series of deep earning models using two different model architectures (i.e., single-stage and dual-stage) to detect shell rings, a circular midden feature that is found across the American Southeast. Native American groups constructed these features during the mid-Holocene (5000-3000 cal B.P.). These deposits offer important information about pre-European contact socioeconomic organization among Native American groups (Figure 1). In the coastal area of the Atlantic, however, these features are relatively rare: only about 50 shell rings have been documented by archaeologists to date. To expand our knowledge of these features, we test RetinaNet and MaskRCNN deep learning models as means of detecting shell rings from wide-area LiDAR data using extremely small training datasets. We demonstrate that the use of “negative” training data to identify non-archaeological features helps to improve model performance. Furthermore, we show that while popular dual-stage detectors like Mask R-CNN perform better than single-stage models like RetinaNet, single-stage models still achieve acceptable levels of accuracy and require a fraction of the computational and time requirements of dual-stage detectors