AI & Archaeology · Holocene · Southern edge of the Nefud Desert, Northern Saudi Arabia
How AI learned to map ancient stone cairns in Saudi Arabia's desert
Three deep learning models scanned satellite images of about 2,500 square kilometres on the southern edge of the Nefud, where Holocene communities built cairns and enclosures.
Peer-reviewed paper
Southern edge of the Nefud Desert, Northern Saudi Arabia A team led by Amy Hatton has trained artificial intelligence to find ancient stone structures on satellite images of northern Saudi Arabia. The study covered about 2,500 square kilometres along the southern edge of the Nefud Desert. The structures include burial cairns and enclosures that are thousands of years old.
The work is published in the Journal of Archaeological Science: Reports. It tests whether computer models can take over part of a slow job: recording these sites one by one. HeritageDaily reports that the models can identify large numbers of structures. The report does not give a count.
Cairns, camps and enclosures in a once wetter desert
People across Arabia built many dry-stone features during the Holocene, the warmer period that began at the end of the last Ice Age, about 11,700 years ago. They include burial cairns, enclosures and camps, as well as larger ritual structures. Archaeologists use them to work out how communities lived and moved as their environment changed.
The study area is dry today and covered by sand dunes. In the past, conditions were wetter. That made the desert edge better suited to settlement, which helps explain why it holds so many remains.
Until now, archaeologists have recorded these sites mainly in two ways. They walk the ground in field surveys, or they trace each structure by hand on satellite images. Both methods take a lot of time and labour.
Three models with different strengths
The team trained three deep learning models, called MA-Net, SegFormer and U-Net, on satellite imagery from the region. Each model was tested on how well it could recognise stone features and trace their outlines.
The models did not perform equally. MA-Net produced the strongest overall scores. SegFormer gave more consistent results across different types of structure.
Image quality made a clear difference. When the models were given higher-resolution images, their accuracy rose by about 20 percent. HeritageDaily's report does not say which image resolutions were used. It also does not describe how the models' outlines were compared with structures mapped by archaeologists.
The researchers say the method does not replace fieldwork. Survey on the ground and excavation are still needed to confirm what the models find and to collect detailed evidence. The models narrow the search instead. They point archaeologists to the places most likely to hold structures and cut the time spent scanning images by hand.
Faster mapping could also help protect the sites. Many lie in remote areas that are hard to monitor. A quicker record of what exists makes it easier to track changes where development or natural erosion may damage the remains.
Machine learning is already used in archaeology elsewhere, but it has seldom been applied in Arabia. According to HeritageDaily, the results show the models can handle desert terrain and large sets of images without major problems. As higher-resolution satellite data becomes more widely available, the same approach could be extended across much larger areas.
Sources
- AI tools speed up search for ancient desert structures in Arabia (HeritageDaily)
Paper: Study led by Amy Hatton, Journal of Archaeological Science: Reports (2026), https://doi.org/10.1016/j.jasrep.2026.105734
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…