AI & Archaeology · Ancient Mesopotamia · Abu Ghraib, Baghdad Governorate, Iraq
How 1960s spy satellite photos helped AI find four new archaeological sites in Iraq
Retrained on CORONA images taken before decades of farming and building, a University of Bologna model spotted four new sites west of Baghdad that fieldwork then confirmed.
Abu Ghraib, Iraq A deep learning model trained on American spy satellite photos from the 1960s has found four previously unknown archaeological sites in Iraq. They lie in the Abu Ghraib district, west of Baghdad. Archaeologists then checked the sites on the ground and confirmed them. None had been identified before with traditional methods.
The work comes from computer scientists and archaeologists at the University of Bologna. A. Pistola, V. Orrù, N. Marchetti and M. Roccetti published it in PLOS ONE in August 2025.
The model looks mainly for tells. A tell is an artificial mound built up from the debris of centuries of settlement, as communities rebuilt in the same place with perishable materials. Tells are the most widespread type of ancient settlement evidence on the Mesopotamian floodplain. They also show up well from above.
The study's summary does not say how old the four new sites are.
Old photos of a landscape that has gone
The team turned to old images because the modern landscape has changed so much. The United States' CORONA reconnaissance programme photographed the region between 1960 and 1972. Comparing those black-and-white images with recent ones showed heavy damage from farming, canal digging and spreading towns.
Fieldwork measured the toll. Of the sites studied, 38% had been completely destroyed. Another 23% had lost more than half of their original extent. Only 38% kept more than half of their surface. So the team trained its models on CORONA images, the only source that captures the landscape before extensive modern changes.
Abu Ghraib was chosen partly because of this heavy transformation. Apart from its easternmost edge, explored by Robert McCormick Adams, the district had never been investigated systematically.
The starting point was an earlier Bologna model. It was trained on modern Bing satellite images and on the outlines of 4,934 surveyed and published sites in the southern Mesopotamian floodplain. The model labels each pixel of an image, so archaeologists can see which areas it marks as a site.
The earlier version had two weaknesses, the authors write. A first attempt to add CORONA images failed, probably because the team could not combine the grey images with colour ones. And it was tested only on hundreds of sites that were already known, so its predictions were never checked on new ground.
Checked against known sites and in the field
For the new study, the team retrained the network and fine-tuned it in two stages. They produced three versions: one using only Bing images, one using only CORONA images, and one combining both.
They checked the results in two phases. First they tested the models on known sites and on areas with no archaeological remains. Then they took the models' new predictions into the field. A survey in 2023 and 2024 visited every possible site flagged by remote sensing, whether spotted by conventional analysis or by the AI.
With CORONA images, performance in the Abu Ghraib area rose. A standard score for how well the model's site outlines match the mapped ones passed 0.85, against 0.81 for the earlier model. Overall accuracy in detecting sites reached 90%, compared with about 80% before. The earlier figures came from a test on already known sites.
The authors also note a general limit of deep learning. Such models can draw inferences that are hard, or even impossible, to explain formally. In this project the AI is one tool among several. The others are conventional remote sensing, collecting finds on the ground and spatial analysis.
The fieldwork confirmed whether remains were present and, where needed, refined site boundaries. Marchetti directed the 2022–2024 Abu Ghraib survey project, which Iraq's State Board of Antiquities and Heritage authorised. The University of Bologna and the Italian Ministry of Foreign Affairs and International Cooperation funded it. The modelling was supported by the KALAM project, funded by the Volkswagen Foundation.
The team used open-source software, public images and publicly available Google Colab computing. All the code, data and archaeological annotations are available on GitHub.
Sources
Paper: Pistola A, Orrù V, Marchetti N, Roccetti M (2025) AI-ming backwards: Vanishing archaeological landscapes in Mesopotamia and automatic detection of sites on CORONA imagery. PLoS One 20(8): e0330419. https://doi.org/10.1371/journal.pone.0330419
Spotted a mistake? Tell the editor about a spelling error or a fact that is wrong. We correct stories openly.
Readers' picks show on the front page.
Follow @AncientHeraldHQ on X for the day's finds as they come in.
Comments
House rules. Stay on the subject, be kind to other readers and say where your facts come from. No advertising and no personal attacks. The editor reads along and hides comments that break these rules.
Loading comments…