AI & Archaeology · Imagery 2016–2023 (site ages not given) · Afghanistan (1,943 sites), Afghanistan
AI spots looted archaeological sites in Afghanistan from monthly satellite images
Trained on 1,943 Afghan sites, the best model scored 0.926 out of 1 at telling looted sites from intact ones, but the study is not yet peer reviewed.
Not yet peer reviewed
Afghanistan (1,943 sites), Afghanistan A computer model can pick out looted archaeological sites in Afghanistan from monthly satellite mosaics, a new study reports. Its best version scored 0.926 out of 1 on a standard accuracy measure. The study was posted on arXiv, a preprint server, and has not yet been peer reviewed.
The researchers write that looting is a severe risk to cultural heritage. Monitoring thousands of remote sites is difficult in practice. They set out to build a satellite-based pipeline that scales to many sites.
Trained on almost 2,000 Afghan sites
The team assembled a dataset of 1,943 archaeological sites in Afghanistan. Of these, 898 were labelled looted and 1,045 preserved. Each site came with an outline of its footprint and imagery covering several years, from 2016 to 2023.
The images are PlanetScope monthly mosaics. Each pixel covers about 4.7 metres of ground, roughly the length of a car. The abstract calls the dataset curated but does not say who decided which sites counted as looted. It also does not give the ages of the sites.
The team compared two kinds of method. The first used convolutional neural networks (CNNs), image-recognition models that learn directly from raw colour image patches. The second fed hand-picked measurements of colour and texture into traditional machine-learning tools.
Other versions of the second approach used summaries produced by recent remote-sensing foundation models instead. These are large models pretrained on satellite data, built to be reused for many tasks.
Pretrained image networks did best
The CNNs won clearly. The best ones were first trained on ImageNet, a large collection of everyday photographs. They also used the site outlines to focus on each site, a step the authors call spatial masking.
These networks reached an F1 score of 0.926. The F1 score balances missed looted sites against false alarms, and 1 would be a perfect result. The best traditional set-up reached 0.710.
That traditional set-up combined location and image summaries from a model called SatCLIP. It fed them into a Random Forest, a standard type of classifier, and averaged the data over time.
Further tests removed one ingredient at a time. They showed that both the ImageNet pretraining and the spatial masking improved results. The pretraining helped even though everyday photos look very different from satellite images, a gap the researchers call domain shift.
The foundation-model summaries performed about as well as the handcrafted features. The authors suggest this is because the signs of looting are extremely localised.
What the study leaves open
A score of 0.926 still means the model makes mistakes. The abstract does not say how many looted sites were missed or how many intact sites were wrongly flagged. It reports results for Afghanistan only.
The team has made its code available in a public repository linked from the paper.
Sources
Paper: Satellite-Based Detection of Looted Archaeological Sites Using Machine Learning. arXiv preprint, 2026. DOI: 10.48550/arXiv.2602.19608
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…