Researchers Use AI and Machine Learning to Reveal Undiscovered Shell Rings

Shell rings and shell mounds left by the indigenous people around 3000 to 3500 years ago are still found in the deep, dense coastal forests and marshes of the American Southeast.

Shell rings located on Daws Island, South Carolina. Both rings are approximately 150 to 200 feet in diameter and are comprised largely of oyster, mussel, and clam shells. Image Credit: Dylan Davis, The Pennsylvania State University

A team of international researchers has now located previously undiscovered shell rings by employing deep machine learning to evaluate remote sensing data. The researchers believe that this will offer better insights into how people lived in that area and find ways to trace other undiscovered shell rings.

The rings themselves are a treasure trove for archaeologists. Excavations done at some shell rings have uncovered some of the best preservation of animal bones, teeth and other artifacts.

Dylan S. Davis, Doctoral Candidate in Anthropology, The Pennsylvania State University

Davis notes that shell rings are considered to be the center of goods exchange. They can offer a lot of data on social constructs, foraging, and politics as well as pointing out the kind of resources that were exploited to identify whether or not they were sustainably utilized. 

The shell rings have produced copper that came from the Great Lakes region to the Southeast. Archaeologists also find ceramics, decorative items and lithics that may have come from up to 100 miles away” added Davis.

In Davis’s perspective, the environments accommodating these shell rings are sometimes challenging to survey. This is because a person standing 2 feet away from the site might not be able to see it.

Instead of observing from the ground level, the scientists utilized three types of available data obtained either from satellite or aircraft—SAR, lidar, and multispectral data. The study has been published in the Journal of Archaeological Science.

The researchers started with the lidar dataset of the southeastern U.S. coast generated by the U.S. National Oceanic and Atmospheric Administration. The datasets related to both East and West coasts of the country are available for the public. Normally, lidar is obtained by a drone or airplane and employs pulses of light to map the surface of an area, and can “see” through forests and also other ground covers.

The scientists utilized a “deep learning” process to teach a Convolutional Neural Network, a kind of neural network employed to test visual information, to recognize shell mounds, shell rings, and other landscape objects.

The team manually reviewed the lidar maps and identified the familiar shell rings. A few of the known rings were reserved to be tested by CNN. They “taught” the neural network with these known rings, using images of mounds and with advanced structures with similar profiles. The researchers also captured images of the known rings and developed more data by turning the image by 45°. Any sites that were changed were also included. 

There are only about 50 known shell ring sites in the Southeastern U.S. So, we needed more locations for training,” stated Davis.

More information was added with the help of SAR data from Synthetic Aperture Radar, belonging to the European Space Agency’s (ESA) Sentimel-1 satellite and multispectral data, imaging beyond the visual spectrum, from ESA’s Sentimel-2 satellite.

SAR can image somewhat through the trees and bushes and can provide information on soil attributes. Multispectral imaging can expose aspects that cannot be seen using human eyes.

Merging these three datasets and utilizing deep training, the researchers could identify hundreds of new shell ring sites, including three to five new shell rings sites in counties where these were never discovered before. The study covered an area spanning three counties, almost 4000 square miles.

Archaeologists are using more and more AI and automation techniques. It can be extremely complicated and requires specific skill sets and usually requires large amounts of data.

Dylan S. Davis, Doctoral Candidate in Anthropology, The Pennsylvania State University

The scientists highlighted that they employed artificial intelligence algorithms already used in ARCGIS, a commercially available geographic information system program. Furthermore, the researchers presented the code and models in the paper so that others can try this type of testing. 

One difficulty with deep learning is that it usually requires massive amounts of information for training, which we dont have when looking for shell rings. However, by augmenting our data and by using synthetic data, we were able to get good results, although, because of COVID-19, we have not been able to check our new shell rings on the ground.

Dylan S. Davis, Doctoral Candidate in Anthropology, The Pennsylvania State University

The research team comprises Gino Caspari, a postdoctoral fellow at the Swiss National Science Foundation; Carl P. Lipo, professor of anthropology and associate dean for research and programs at Binghamton University; and Matthew C. Sanger, curator at the National Museum of the American Indian.

The study was financially supported in part by the Swiss National Science Foundation.

Journal Reference:

Davis, D. S., et al. (2021) Deep learning reveals extent of Archaic Native American shell-ring building practices. Journal of Archaeological Science.


Tell Us What You Think

Do you have a review, update or anything you would like to add to this news story?

Leave your feedback