Sunday 15 November 2026SearchSign in
Ancient Herald

News from the ancient world

AI & ArchaeologyBronze Age · Porto Torres, Olmedo and Stintino, Sardinia, Italy

An AI trained on Google Maps finds Bronze Age nuraghe towers in Sardinia, and possibly six new ones

The AIOH model learned from 1,612 satellite images of known nuraghes. A first field check in north-west Sardinia found six probable unrecorded sites, and eight false alarms.

Stone Bronze Age tower complex of Su Nuraxi in Sardinia
Su Nuraxi at Barumini, the best-known nuraghe complex in Sardinia, with its central tower and surrounding village. Photo: Norbert Nagel / CC BY-SA 3.0

In brief 3 key points

ShowHide
  • An AI model spots Sardinia's Bronze Age stone towers in satellite images
  • A quick field check found six probable sites not in any record
  • Eight flagged spots were false alarms, such as rock-covered landfill

An artificial intelligence model trained on Google Maps images can pick out Sardinia's Bronze Age nuraghes, the tower-shaped stone monuments found only on the island. In a first, quick field check, the team behind it found six probable archaeological sites that were not in any record. The study appears in the journal Archeologia e Calcolatori.

The system is called AIOH, short for Artificial Intelligence Operator for Heritage. It was built by Augusto Palombini and colleagues, with the CNR Institute of Heritage Science in Italy. The goal was a tool that could scan whole regions using freely available satellite images.

Dry-stone Bronze Age tower of Nuraghe Losa
Nuraghe Losa near Abbasanta, a Bronze Age tower with added bastions, in central Sardinia Photo: CV1958 / CC BY 4.0

Five months of labelling

Nuraghes vary in size and shape, and many are hard to see from above. Some stand in quarries, woods or towns, some hide under trees and some have been swallowed by later buildings. Labelling the training images by hand took about five months.

The final dataset, called NuragAI, held 1,612 images with a visible or hardly visible nuraghe and 6,751 images without one. An archaeologist checked the empty images. The team added round modern buildings and water towers so the model would learn not to confuse them with nuraghes.

Interior of a round dry-stone hut
Inside a round stone hut of the nuragic village around Su Nuraxi, Barumini Photo: Norbert Nagel / CC BY-SA 3.0

The model is a vision transformer, a type of neural network that looks at the whole image at once. On the test images it scored almost perfectly, with an average precision of 99.93 per cent. The authors write that such high results call for caution, because the two sets of images may be too uniform.

Out in the landscape

In a 60 by 30 km area of north-western Sardinia, the model gave each grid cell a probability of holding a nuraghe. Results there were more moderate. Some known monuments fell in low-probability cells, often because they are poorly preserved or their positions differ between records.

Every precisely located nuraghe in the three test municipalities, Olmedo, Stintino and Porto Torres, lay within 250 m of a cell rated above 80 per cent. Heatmaps showed the model also weighs the surrounding landscape. The authors suggest it may detect a nuragic area, ground favoured by nuraghe builders, rather than only the towers.

Tall stone tower of Nuraghe Is Paras
The central tower of Nuraghe Is Paras at Isili, built around 1500 BC Photo: Norbert Nagel / CC BY-SA 3.0

The team then visited 24 high-scoring spots, without permits to dig or enter private land. They rated four sites as probable and two more as probable from a distance, with traces such as pottery on the ground and hammer marks on rocks. Three were possible, five could not be checked and ten were negative, two of them close to known archaeological traces.

The false alarms included a pile of landfill covered with rocks and the dry-stone fence of a recent garden. None of the six probable sites appears in the regional archaeological maps. The authors call the test preliminary and say talks with local authorities are under way.

Sources (2)

Peer-reviewed paperfrom CNR Institute of Heritage Science

Paper: Palombini, A., Baiocchi, E., Malatesta, S. G., Lanzino, R., Marini, M. R., Rosati, P., AIOH: Artificial Intelligence Operator for Heritage. The case study of nuraghe detection, Archeologia e Calcolatori 37.1 (2026), 229-254, https://doi.org/10.19282/ac.37.1.2026.12

Edited by Guiseppe Weijermars

Spotted a mistake? Tell the editor about a spelling error or a fact that is wrong. We correct stories openly.

nuraghesardiniabronze ageartificial intelligencesatellite images

The week in archaeology

The week's best finds in one e-mail, every Friday. Free, no ads.

Unsubscribe with one click. See the latest issue · Privacy

Related stories

Most read this week

  1. 1
    Iron AgeAncient DNA of 4,587 people points to central Europe as the cradle of the Celtic languages
  2. 2
    PrehistoryBlack Sea shells found among 41,000-year-old jewellery at Russia's Kostenki
  3. 3
    PrehistoryA 75,000-year-old Neanderthal from Shanidar Cave was killed by a wooden spear, and had been cared for after an earlier injury
  4. 4
    Iron AgeVerica, the British ruler who brought Claudius to invade, was a woman, Bournemouth archaeologist argues
  5. 5
    PrehistoryTaung Child, the fossil that put human origins in Africa, is about 3 million years old