Using deep learning to detect rare archaeological features: A case from coastal South Carolina, USA Dylan S. Davis* – Department of Anthropology, The Pennsylvania State University Gino Caspari – Department of Archaeology, University of Sydney Institute of Archaeological Sciences, University of Bern Matthew C. Sanger – National Museum of the American Indian, Smithsonian Institution Carl P. Lipo – Department of Anthropology, Binghamton University Summary 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 learning 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 Mask- RCNN 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. Introduction Shell rings are circular refuse piles composed of plant and animal remains that surround an empty central plaza (Russo 2004; Sanger 2017). These deposits represent some of the earliest evidence of permanent human occupation in the coastal regions of the American Southeast (Figure 1; Russo 2004). The nature of past community activities that produced these shell-rings remains debated among archaeologists, some of whom have focused on identifying the degree to which these deposits served as residential, ritualistic, or a mixture of mundane and ceremonial activities (Russo 2004; Sanger and Ogden 2018; Trinkley 1985). Despite continued investigation, current archaeological knowledge of shell ring distribution is limited, as only about 50 have been recorded in the entirety of the region (Figure 1). Because of their location is often within difficult-to- survey, dense forests and marshlands, these deposits are mostly known on the basis of large examples that are the most recognizable and most easily accessed (Davis et al. 2020). As a result, archaeologists lack a full inventory of extant shell rings that would allow for a comprehensive investigation into the morphological variability of shell rings and the range of contexts in which they are found. Deep learning, a branch of machine learning, has been rapidly gaining popularity among archaeologists seeking to identify features among large remote sensing datasets in the past several years (e.g., Caspari and Crespo 2019; Trier, Reksten, and Løseth 2021). Convolutional Neural Networks (CNNs), in particular, have proven highly effective at increasing true positives while reducing false positive results in object detection studies (e.g., Caspari and Crespo 2019; Lambers, Verschoof-van der Vaart, and Bourgeois 2019). Yet, applications of deep learning within archaeology have been limited because of the amount of training data required Figure 1: Location of confirmed shell rings surrounding the study area (black box). Service Layer Credits: ESRI, HERE, GARMIN, OpenStreetMap contributions, and the GIS User Community 10.1190/segam2021-3574963.1 Page 3265 © 2021 Society of Exploration Geophysicists First International Meeting for Applied Geoscience & Energy Downloaded 09/22/21 to 34.228.166.90. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms DOI:10.1190/segam2021-3574963.1