Japanese scientists from the University of Yamagata have discovered four new geoglyphs in the arid Peruvian coastal plain, in the northern part of the Peruvian Nazca Plateau, using deep AI training. The research has been conducted since 2004 by a team from the University of Yamagata led by Professor Makato Sakai. The University of Yamagata is conducting geoglyph distribution studies using satellite imagery, aerial photography, onboard LiDAR scanning, and drone photography to explore the vast area of the Nazca Plateau, covering over 390 km2.
The Nazca Lines, according to scientists, were created over centuries, beginning around 100 B.C., by the Nazca people of present-day Peru. They were first studied in detail in the 1940s, and by the time they were added to the UNESCO World Heritage List in 1994, about 30 had been identified. They are remarkably well preserved given their age.
Archaeologists discovered 142 new drawings in the desert over the course of a decade, identifying them by hand with aerial and on-site photography. Then, in collaboration with researchers from IBM Japan, they used machine learning to search the data for designs that had been missed in previous studies.
The four new geoglyphs depict a humanoid figure, a pair of legs, a fish and a bird. The humanoid geoglyph holds a club in its right hand and is 5 meters long. Geoglyph of a fish with its mouth wide open is 19 meters long, geoglyph of a bird is 17 meters, and a pair of legs is 78 meters long.
A study published in the Journal of Archaeological Science showed the discovery of four new Nazca geoglyphs using this new method by developing a training data labeling approach that identifies a similar partial pattern between known and new geoglyphs.
“We have developed a deep learning system that solves problems often encountered in the task of discovering archaeological image objects,” the authors of the study write.
These results serve as another illustration of how machine learning can be useful to scientists, especially for tasks involving significant datasets. Like humans, algorithms can be taught to sift through certain types of data in search of patterns and anomalies. Although creating these tools can be challenging, once trained, such algorithms become indefatigable and consistent.