Deep learning reveals extent of Archaic Native American shell-ring building practices. (August 2021)

• Record Type: Journal Article
• Authors: Davis, Dylan S.; Caspari, Gino; Lipo, Carl P.; Sanger, Matthew C.
• Is Part Of: Journal of archaeological science. Volume 132(2021)
• Publication Date: 2021-08
• Subjects: Settlement patterns -- Deep learning -- Shell rings -- Object detection -- Southeast United States; Archaeology -- Periodicals; Archéologie -- Périodiques; 930.1
• DOI: 10.1016/j.jas.2021.105433
• Languages: English
• ISSNs: 0305-4403

Abstract

Abstract: In the mid-Holocene (5000 - 3000 cal B.P.), Native American groups constructed shell rings, a type of circular midden, in coastal areas of the American Southeast. These deposits provide important insights into Native American socioeconomic organization but are also quite rare: only about 50 such rings have been documented to date. Recent work using automated LiDAR analysis demonstrates that many more shell rings likely exist than are currently recorded in state archaeological databases. Here, we use deep learning, a form of machine intelligence, to detect shell ring deposits and identify their geographic range in LiDAR data from South Carolina. We corroborate our results using synthetic aperture radar (SAR), multispectral data, and a random forest analysis. We conclude that a greater number of shell rings exist and that their distribution expanded further north than currently documented. Our evidence suggests that ring-construction was a more widespread and common practice during the mid-Holocene. Highlights: Shell ring architecture expands ~50 km further north than previously documented. Deep learning can systematically explore spatial extents of cultural practices. Mask R-CNN can detect rare archaeological features with small training datasets. Using multiple datasets and analytical methods improves site detection efforts. Shell ring building appears to be a common practice during the Archaic Period.