AI model helps detect Iraqi archaeological sites
A team of computer scientists and archaeologists from the University of Bologna in Italy has developed a new tool for identifying archaeological sites using artificial intelligence (AI). Publishing their results in the journal Nature Scientific Reports, the team reached a predictive accuracy of over 80 percent, potentially dramatically increasing the rate at which archaeologists can identify, survey, and preserve previously unknown sites.
The Italian team trained a deep-learning computer program on the thousands of archaeological sites found in the floodplains of southern Iraq. Using satellite images, declassified Cold War satellite imagery, and data from nearly 5,000 known archaeological sites across an area roughly the size of West Virginia, the team trained the program to pick out tells (archaeological mounds, see below) from the rest of the human and natural landscape of southern Iraq. The resulting model “predicted” potential archaeological tells with an accuracy of over 80 percent.
Traditionally, identifying archaeological tells was done through ground and aerial surveys, while in recent years, remote sensing using satellite imagery, LiDAR, and other high-tech methods has become more common. However, all of these methods require extensive time and work from highly trained archaeologists. While not without its flaws, the AI program drastically cut down on the amount of time needed to identify potential sites, thereby allowing archaeologists to focus their attention on just the most likely candidates. Additionally, since the program uses satellite imagery available through sources like Google Maps, it is extremely cost-effective.
However, AI is still not archaeology’s instant fix to locating new sites. With the model’s 80 percent accuracy, archaeologists still need to “ground truth” the results through surface survey. Additionally, the model is not applicable to all types of sites. Many smaller tells cannot be identified because they easily blend in to the local environment, while in other cases, later human activity and natural processes have degraded or even leveled entire sites, making them impossible to detect without a ground survey.
The program is also not yet capable of detecting sites in different types of regions and landscapes. Testing the AI program on satellite images from Uzbekistan, for example, the team noted that the model’s predictive was only 20 percent. This was likely a result of the very different environment in Uzbekistan and the different types of sites and settlement patterns. By providing new training sets for such regions, however, the researchers believe it would be possible to achieve much better results.
A tell, as described by Oded Borowski, Professor Emeritus of Biblical Archaeology at Emory University, in his article “How to Tell a Tell,” is:
A succession of cities built on top of one another. Usually, the first occupation of the site was on a low hill overlooking the surrounding area. Each successive city was built on the ruins of the previous one. Each city came to an end, however, destroyed by an invading army or, sometimes, abandoned because of changing climatic or political conditions. Each city left a layer of deposit on top of which the next city was built. The result of this build-up resembles a multi-layered chocolate cake, each layer representing a destroyed or abandoned city.
Tell sites often have very characteristic features that can make them stick out like a sore thumb to trained archaeologists. However, whether because of environmental changes, human causes, or even a site’s secluded nature, many of them go completely unnoticed. With growing worries about climate change and other processes that can damage sites, it is more important than ever to find and document them. Hopefully, new methods like the one proposed by the Italian team will make this all the easier.
Sign up to receive our email newsletter and never miss an update.