Semi-automated detection of Holocene archaeological structures along the southern edge of the Nefud desert

Bibliographic Data

ID19488607
AuthorsAmy Hatton (0000-0002-2900-3681, University of Cape Town, corresponding author), Amina Jambajantsan (Max Planck Institute for the Science of Human History), Paul S Breeze (0000-0003-2575-8021, King's College London), Maria Guagnin (0000-0002-7995-8611, The University of Sydney), Michael T Fisher (0000-0003-2865-210X, Max Planck Institute for the Science of Human History), Faisal al-Jibreen, Abdullah M Alsharekh (0000-0002-2666-3262, King Saud University), Michael D Petraglia (0000-0001-5434-6793), H S Groucutt (0000-0002-9111-1720, University of Cologne)
Year2026
Volume72
Pages105734
Publication date2026-06-01
Peer ReviewedYes
Open AccessYes
TypeARTICLE
VenueJournal of Archaeological Science Reports (JOURNAL)
Journal identifiersISSN: 2352-409X • E-ISSN: 2352-4103
PublisherElsevier BV (PUBLISHER)
DOI10.1016/j.jasrep.2026.105734
OpenAlexW7153194007
LanguageEN
References cited47

Throughout the Holocene, humans covered the landscape of Arabia in hundreds of thousands of diverse dry-stone structures. These archaeological remains are predominantly funerary structures, enclosures, and camps although they also include ritualised features such as mustatils , all of which form essential components for understanding Neolithic and Bronze Age peninsular societies and their dynamic relationships with the natural environment. Archaeologists typically document these structures by either mapping them in the field or applying manual digitisation within a Geographic Information System (GIS) framework. Such methods are time consuming and can be expensive, especially in the case of field documentation. In order to develop a more time- and cost-effective method for identifying large numbers of stone features, we have tested three pre-trained semantic segmentation Deep Learning models (MA-Net, SegFormer, and U-Net) for semi-automatic feature detection by applying them to satellite imagery of archaeological landscapes in northwestern Saudi Arabia. The results show that, the MA-Net model performed best on averaged metrics, however, the SegFormer model showed more stable metrics during training. We present confusion matrices to show that the SegFormer model is more consistently able to correctly identify stone structures, regardless of structure types, in contrast to the U-Net and MA-Net models. While survey and excavation of these structures is essential for producing fine-resolution data, automated workflows that incorporate remote sensing data can generate the breadth of coverage required for interpreting ancient social landscapes on a wider geographic scale. Such mapping is also critical for defining and protecting cultural heritage across the vast arid landscapes of desert regions such as Arabia and the Sahara.

Assemblage (archaeology) · Desert (philosophy) · Enhanced Data Rates for GSM Evolution · Holocene · Radiocarbon dating · Archaeological Research and Protection · Archaeology and Historical Studies · Image Processing and 3D Reconstruction

Tools

Open DOIOpen Access