Mapping the stones in an ancient wall by hand is grueling work. An archaeologist crouches against sun-baked masonry, sketches each block, measures its edges with calipers or a total station, and repeats the process hundreds of times before a single wall section is fully documented. A team at the University of Haifa believes it has built something that can do much of that job from the air.
In a paper published in May 2026 in the Journal of Archaeological Science (volume 185, article 106430), researchers Erel Uziel, Motti Zohar, and Yitzchak Jaffe describe a semi-automatic AI system that takes high-resolution drone photographs of archaeological ruins and segments them into individual building stones and wall sections. According to the authors, the system achieves sub-centimeter accuracy, a level of detail that, if independently validated, would rival the precision of ground-based laser scanning at a fraction of the time and cost.
What the system does
The pipeline starts with orthomosaics: large, geometrically corrected overhead images stitched together from overlapping drone photographs. These composites flatten perspective distortion so that measurements taken from the image correspond reliably to real-world distances. The Haifa team’s machine learning model then analyzes these images pixel by pixel, drawing boundaries around each stone and grouping stones into wall segments.
What makes this harder than it sounds is the nature of ancient masonry. Stones in a wall are tightly packed, partially hidden by mortar or vegetation, and wildly inconsistent in shape, color, and surface texture. Shadows cast by uneven surfaces can trick algorithms into seeing edges that aren’t there or missing ones that are. Earlier work in the same journal demonstrated that machine learning could detect broken pottery fragments from drone imagery, but identifying a potsherd sitting on open ground is a fundamentally simpler problem than distinguishing one limestone ashlar from its neighbor inside a mortared wall.
The Haifa team describes their tool as “semi-automatic,” which means a human operator remains in the loop. The exact division of labor between algorithm and archaeologist is not fully detailed in the publicly available metadata, but the term implies that users review, and likely correct, the AI’s output rather than accepting it blindly.
Open code, open models
Perhaps the most striking aspect of this release is its transparency. Alongside the paper, Uziel, Zohar, and Jaffe deposited a full reproducibility package on Zenodo, the open-access repository operated by CERN. The package includes pre-trained models, sample orthomosaics, configuration files, and step-by-step instructions. Any research team with a reasonably powerful computer and its own drone imagery can download the package and run the pipeline without building a detection system from scratch.
That level of openness is still uncommon in archaeological AI, where code and training data are frequently kept private or shared only upon request. By assigning the deposit its own DOI (10.5281/zenodo.15673667) through DataCite, the authors have created a stable, citable entry point that other researchers can reference, test against, and build upon. The inclusion of pre-trained weights signals that this is not a proof-of-concept script but a tool the team intends others to use immediately.
The sub-centimeter question
The claim of sub-centimeter accuracy deserves careful scrutiny. A drone image can have a ground sample distance (GSD) of just a few millimeters, meaning each pixel represents a tiny patch of ground. But pixel resolution and geometric fidelity are not the same thing. Lens distortion, slight errors in camera positioning, and the irregular surfaces of ancient walls all introduce uncertainty between what the image shows and where a stone actually sits in three-dimensional space.
A dataset paper published in Scientific Data by Nature Portfolio, which used laser scanning to create digital twins of three-leaf stone walls, draws exactly this distinction. That study established rigorous benchmarks for capturing individual stone geometry in masonry and serves as a useful yardstick for evaluating any new stone-mapping tool. Until independent teams replicate the Haifa pipeline on their own sites and compare the results against ground-truth measurements from total stations or terrestrial laser scanners, the sub-centimeter claim rests on the authors’ own reported metrics.
Specific performance figures, including detection precision, recall rates, and measured positional error, are detailed in the full journal article. The publicly available bibliographic records and Zenodo metadata confirm the system’s purpose and structure but do not reproduce those numbers, so independent verification will depend on access to the complete paper and on replication studies by other groups.
What we don’t know yet
Several important questions remain open. The specific archaeological sites used to train and test the model have not been named in the accessible records. That matters because stone construction techniques, materials, and weathering patterns vary enormously across regions and historical periods. A model trained on Hellenistic limestone walls in the Levant might struggle with medieval basalt fortifications in Turkey or mudbrick foundations in Mesopotamia without retraining.
The exact deep learning architecture, whether the team adapted a standard instance segmentation framework like Mask R-CNN or built a custom network, is likewise not specified in the metadata. Nor is it clear how large the training dataset was or how many wall types it covered. These details, almost certainly present in the full paper, will determine how broadly the tool can be applied out of the box.
Quantitative comparisons to traditional survey methods are also absent from the sources reviewed here. Field archaeologists will want to know how many hours the AI pipeline saves relative to manual photogrammetric mapping, and whether the output meets the accuracy thresholds required for conservation-grade documentation, structural monitoring, or heritage protection filings. The “semi-automatic” label suggests meaningful efficiency gains, but the magnitude of those gains has not yet been independently reported.
Where this fits in the field
Drone-based survey has already transformed how archaeologists discover and record sites. Thermal imaging reveals buried structures, multispectral cameras highlight crop marks that trace ancient foundations, and photogrammetry generates detailed 3D terrain models. What has been missing is the ability to move from broad site-level mapping down to the granular scale of individual construction elements without putting boots on the ground and tape measures on the wall.
The Haifa system targets that gap. If its accuracy holds up under independent testing, it could accelerate documentation at large, stone-built sites where hundreds of wall segments need recording, places like Caesarea, Megiddo, or Jerash where excavation seasons are short and budgets are tight. It could also prove valuable for monitoring structural change over time: run the pipeline on drone imagery captured six months apart, and any shifted or missing stones should show up in the comparison.
For now, the strongest conclusion supported by the evidence is that a real, reproducible tool exists, it is openly available, and it is designed to solve a genuine bottleneck in archaeological fieldwork. Practitioners interested in experimenting with it can access the full package on Zenodo today. Those considering it for conservation-grade work should pair it with local accuracy checks and cross-comparisons against established survey methods until the broader community has had time to put the system through its paces.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
