An archeologist and a computer scientist have developed a new technique to analyze satellite images of wide swaths of land to discover ancient settlements in the Middle East. In their PNAS paper, “Mapping patterns of long-term settlement in Northern Mesopotamia at a large scale,” Harvard Associate Professor of the Social Sciences Jason Ur and Bjoern H. Menze, a research scientist in the Computer Vision Laboratory at ETH Zurich and a research affiliate of the Medical Vision group of MIT’s CSAIL, describe 8,000 year-old mounds, or “tells,” and 3,000 year-old paths traversed between them in ancient Mesopotamia (now Iraq, northeast Syria, southeast Turkey, and southwest Iran).
The method, which uses machine learning, focuses on the type of soil resulting from long-term human activity. Because it is richer in organic matter than its surroundings, the dirt refracts spectra differently, revealing large and small tells. Their elevations and volumes in turn provide clues to the sizes of homes and populations.
As a graduate student at the University of Heidelberg, Menze used machine learning to evaluate spectroscopic MRI images of cancer patients who had been treated with radiation therapy. His task was to analyze spectra that were likely to have been refracted by brain tumors. Detecting subtle changes in spectra is tedious, however, so clinicians asked him to automate the process.
While doing so, Menze attended several archeology classes. At one lecture, Oxford archeologist Andrew Sherratt showed data downloaded from the Internet, before Google Earth, revealing tells in the Near East. “He talked about how great it would be to record them because they all show up in digital elevation models,” says Menze. He started using those models to train a classifier and found sites in a large landscape in Syria’s Kahbur Basin. Eventually he and Ur extended the analysis to combine the signatures of mounds with the signatures of the soil in spectral images from the ASTER sensor aboard NASA’s Terra satellite.
They identified a signal in the spectral images consistent for a few well-known archaeological sites and then averaged out noise, like short-term variations due to crops and changing weather conditions. The researchers accomplished this by looking not at a single image, but at all images they had of the Basin, some 40 images from a 10-year period.
They transformed the problem into a binary classification task, categorizing a location as a settlement or non-settlement. Next they trained the random forest algorithm, which was developed a decade ago by Leo Breiman of the University of California, Berkeley, to discriminate between the two classes and recognize spatial patterns. “The algorithm works well for high-dimensional spaces,” says Menze. “It’s robust, and trains fast.”
With images from the Shuttle Radar Topography Mission, the researchers also measured, for the first time, the volume of settlement mounds. Volume results from dwellings, like mud-brick houses a family might have occupied for 30 years that was leveled by grandchildren, who built another house on top, for example. They yield a rough estimate of the population that lived in the Basin over the centuries. Ten million people occupied 14,000 sites within 23,000 square kilometers during 8,000 years, according to the researchers. “This was most astonishing to me,” says Menze. “That we found that many sites and could measure indirect population indicators over time, and that the number is as big as the population of modern Syria.”
Ancient travel routes between tells also yielded surprises. Geographic centrality did not determine the average number of visits to a site, but rather the volume of the site did, according to Menze and Ur. Large tells dominate the transportation network for both local or long distance exchanges. “This … suggests that patterns of movement would have been as stable as patterns of settlement,” the researchers write.
Menze was delighted in learning more about the random forest algorithm by using it for interdisciplinary work. “It’s interesting to me to bring one part of computer science to another because in machine learning, things are data driven. The tumor growth model I work with is theoretical and we could not use clinical data for specific problems. But if I have a second problem, like an archaeological application that I can apply algorithms to, that’s great fun.”
Besides clinical applications, these techniques may be useful to industry. Menze and Ur have been contacted by mining companies interested in soils and minerals and short-term variations versus long-term patterns in the soil and minerals on the ground.
Karen A. Frenkel is a science and technology journalist located in New York City.
No entries found