Machine Learning Reads 235,000 Stones to Map an Ancient Town's Growth
Image: "Shivta 101025 Stable House 03" by Oren Rozen (CC BY-SA 4.0) via Wikimedia Commons, recoloured.
Can machine learning turn scattered rubble into evidence about how a settlement grew? Uziel, Avni, Feinman, Bar-Oz and Jaffe (PNAS, 2026) automatically digitized about 235,000 surface stones at Late Antique Shivta and found masonry patterns that reveal a previously unknown earlier boundary and outward growth from a southern core.
Making the building stone the unit of analysis
Archaeologists have long treated buildings as records of how communities organized space and invested labour. Measuring that record systematically is hard, though: existing methods tend to be labour-intensive, focus on single structures, or rely on rough volume estimates. In a new article in PNAS, Erel Uziel, Gideon Avni, Gary M. Feinman, Guy Bar-Oz and Yitzchak Y. Jaffe propose a different starting point. Instead of the building, they make the individual stone the unit of analysis.
Their machine-learning framework combines two data sources. Low-altitude drone imagery captures surface remains across the site, while close-range photographs document standing walls. The authors applied the method to Shivta, a Late Antique settlement in the Negev Desert, and used it to quantify the size, shape regularity and arrangement of stones. Because each stone is measured, the results can be aggregated at several levels: walls, households, insulae (city blocks) and the settlement as a whole.
According to Uziel, Avni, Feinman, Bar-Oz and Jaffe (PNAS, 2026), the method automatically digitized approximately 235,000 surface stones, plus around 12,000 stones in 604 standing walls. Higher masonry scores clustered around churches and along parts of the settlement boundary. More strikingly, the spatial distribution of masonry variation pointed to an earlier boundary that had not been identified before, along with evidence that the town expanded outward from a southern core.
When the authors combined these patterns with architectural sequencing, a chronological picture emerged: builders used larger and more regularly shaped stones during Shivta's main phase of growth. The study therefore links a fine-grained physical measure to a historical process, settlement expansion, that is usually inferred from far coarser evidence.
The evidence comes from a single site, and the abstract frames the broader claims as potential: the authors describe the framework as flexible and robust and argue that it supports application to other stone-built contexts and cross-site comparison, rather than reporting such comparisons. The approach also complements, rather than replaces, existing methods such as architectural energetics, which estimate the labour invested in construction.
Why it matters: For social scientists, this is a clear example of machine learning as a measurement instrument. It turns a mass of unstructured visual material into a dataset at the level of the smallest meaningful unit, then aggregates upward to answer questions about organization and change. The same logic applies well beyond archaeology: researchers working with images of streetscapes, storefronts, documents or products can ask what their equivalent of the single stone is, and whether measuring it at scale would reveal patterns that case-by-case coding misses.
Uziel, E., Avni, G., Feinman, G. M., Bar-Oz, G., & Jaffe, Y. Y. (2026). Quantifying ancient architecture: Stone-level machine learning reveals architectural variation and settlement growth at Shivta. *Proceedings of the National Academy of Sciences*, 123(44). Read the original. Independent editorial summary of third-party research.
Comments