Daws Island, South Carolina has shell rings. The rings measure approximately 150-200 feet in diameter. They are mainly made up of shell, mussel, and clam shells. Credit: Dylan Davis, Penn State
Shell rings and mounds that were left behind by Indigenous peoples 3,000 to 5,000 year ago are hidden deep in the American Southeast's dense coastal forests and marshes. A team of international researchers has found previously unknown shell rings using deep machine learning. This will allow researchers to better understand the area's history and identify any other undiscovered shell rings.
Dylan S. Davis (Penn State doctoral candidate) said that the rings are an archeologist's treasure trove. "Excavations at shell rings have revealed some of the most remarkable preservations of animal bones, teeth, and other artifacts."
According to Davis, shell rings could be considered centers for the exchange of goods. These rings can give a lot information about social constructs, politics, and foraging. They may show how resources were used and whether or not they were sustainable.
Davis stated that the shell rings produced copper from the Great Lakes region of the Southeast. "Archeologists also discover ceramics, decorative objects and lithics from as far as 100 miles away," said Davis.
The left side shows training examples of shell rings and the right side displays the deep learning identifications. You can see that the computer draws a box around most of the objects to identify known examples for shell ring architecture. Credit: Dylan Davis, Penn State
Davis says that the shell rings can be so hard to see in certain environments, that it is possible to stand within 2 feet of one site and not even notice it.
Instead of looking from the ground, researchers used three types existing data, including data from satellitelidars, aircraft, and multispectral. The Journal of Archaeological Science published the results of their research in its most recent issue.
The U.S. National Oceanic and Atmospheric Administration provided a lidar dataset for the southeastern U.S. Coast. These datasets are available for both the East coast and West coasts. Lidar is usually taken by drone or plane and uses pulses to map an area's surface. It can see through trees and other groundcover.
Researchers used "deep learning" to teach Convolutional Neural Networks, a type of neural network that analyzes visual information to recognize shell rings, mounds, and other landscape objects. They manually reviewed the lidar maps to locate known shell rings. They reserved some known rings for later testing of the CNN. Then they "taught" their neural network with these known rings. Images of mounds were used as well as modern structures with similar profiles. They also took photos of known rings and turned the images 45 degrees to create more data. These sites were also included.
Davis stated that there are only 50 shell ring sites known in the Southeastern U.S. "So, we needed additional locations for training."
LiDAR data shows shell rings. These rings are distinctive because of their elevation and slope in comparison to the surrounding landscape. Credit: Dylan Davis, Penn State
SAR datasynthetic radar (SAR) from the European Space Agency's Sentinel-1 Satellite, and multispectral dataimaging above the visual spectrumfrom ESA’s Sentinel-2 Satellite added to the information. SAR is able to see through trees and shrubs and provide information about soil attributes. Multispectral imaging is a way to see features that are not visible by the human eye.
These three datasets were combined with deep training and the researchers were able potentially to identify hundreds of shell ring locations, including three to five new sites in counties that had never seen these rings before. The area covered approximately 4,000 square miles and included three counties.
Davis stated that archeologists are increasingly using AI and automation techniques. It can be very complicated, requires specific skills and often requires large amounts data.
Researchers note that they used artificially intelligent algorithms already available in ARCGIS (a commercially-available geographic information system program). The code and models are also provided in the paper to allow others to attempt this type analysis in other areas.
Davis stated that deep learning has a problem because it requires a lot of information to train, which is not what we have when searching for shell rings. "However, we were able achieve good results by augmenting our data and using synthetic data. However, COVID-19 has prevented us from being able to verify our new shell rings on ground.
Gino Caspari is a postdoctoral fellow with the Swiss National Science Foundation. Carl P. Lipo is a professor of anthropology at Binghamton University and associate dean of research and programs. Matthew C. Sanger is the curator at the National Museum of the American Indian.
Continue reading Novel compound reveals the fundamental properties of the smallest carbon nanotubes
