A model can put a box around a possible earthwork. It cannot establish the feature’s date, significance or legal status, and it cannot recover the context destroyed by an unnecessary excavation. Archaeological AI is useful when it helps specialists examine more evidence while preserving the chain from source image to interpretation.
That chain is the product: source → processing → candidate → expert interpretation → field verification → record → archive. If any step becomes detached, a dramatic “discovery” may be impossible to reproduce or safely manage.
This guide reflects UK sources available on 31 July 2026. Heritage protection and finds law differ across England, Scotland, Wales and Northern Ireland. Consent, land access, planning and reporting must be checked for the exact location before survey or ground disturbance.
Begin With a Research Question and Evidence Register
Do not start by asking a model to “find archaeology.” Define the feature class, landscape, source data and decision the result will inform. A useful brief might be: “prioritise possible charcoal-platform earthworks for specialist review in this woodland lidar tile,” not “locate hidden settlements.”
Create an evidence register:
| Field | What to record | Why it matters |
|---|---|---|
| Source | Provider, flight/survey date, sensor, resolution, coordinate reference system and licence | Establishes what was actually observed and whether it may be reused |
| Processing | Software, parameters, terrain model, visualisation and transformations | Prevents a processed image being mistaken for raw measurement |
| Candidate | Geometry, class, confidence, model version and alternatives | Preserves uncertainty instead of forcing one label |
| Context | Geology, land use, known records, disturbance and relevant historic maps | Helps separate archaeology from natural or modern features |
| Review | Archaeologist, interpretation, date and reason | Makes professional judgment visible |
| Verification | Field method, permission, result and negative evidence | Closes the loop without erasing non-confirmation |
| Deposit | Historic Environment Record and archive identifiers | Makes the evidence discoverable after the project ends |
Keep negative and rejected candidates. They are essential for evaluating error and for preventing another team from repeating the same unsupported interpretation.
Use Lidar as Measured Terrain, Not a Ready-Made Map
Airborne laser scanning can reveal subtle topography beneath some vegetation and across large landscapes. Historic England’s airborne remote-sensing overview describes how specialists interpret aerial photographs and 3D lidar to map archaeological landscapes. Interpretation remains central.
Terrain-model choice, point density, vegetation removal, interpolation, hillshade direction and visualisation can reveal or hide different features. Keep the original point cloud or source product and produce multiple documented views. A linear bank should persist across sensible processing choices; an artefact that appears under one extreme filter needs caution.
Historic England’s June 2026 standards for Aerial Investigation and Mapping require a systematic, integrated approach to aerial sources, mapping, monument recording and GIS outputs compatible with Historic Environment Records. Use those standards as the project frame; add machine learning inside it rather than replacing it.
For wide-area context, satellite teams can apply similar evidence controls from our AI earth-observation guide.
Evaluate Candidates by Landscape and Consequence
A random hold-out of image chips can leak the same site, survey campaign or terrain into training and test sets. Split evaluation geographically and, where possible, by acquisition period. Test the conditions the model will meet: woodland, arable cropmarks, upland, urban fringe, quarrying, drainage, forestry tracks, variable point density and modern earthworks.
Measure:
- precision and recall by feature class and landscape;
- false candidates per square kilometre and the expert time they consume;
- detection of known features withheld from training;
- performance under different source resolutions and processing;
- agreement and disagreement between reviewers;
- percentage of outputs marked “uncertain” or “not assessable”; and
- whether field verification changes the interpretation.
Set the threshold by consequence. A broad research prospection layer can tolerate more candidates than a planning screen that might direct intrusive work or imply a protected site is absent. “No model detection” must never be translated into “no archaeology.”
Put Consent Before Ground Truth
Field verification may involve access, drone operation, geophysics, metal detection, coring or excavation. Model confidence grants none of those permissions.
The government’s scheduled monument consent service explains that work affecting a scheduled monument requires prior consent and that using a metal detector or removing a detector find from a protected site without the relevant licence is an offence. England, Scotland and Wales have different responsible heritage bodies; Northern Ireland has its own consent route.
Build a pre-field gate that checks:
- landowner and occupier access;
- designation and Historic Environment Record information;
- planning, scheduled-monument and environmental constraints;
- survey-specific permissions and safe system of work;
- finds ownership, reporting and archive arrangements; and
- a method proportionate to the research question.
Prefer the least destructive method that can answer the question. A model-generated point of interest is not justification for digging.
For drone survey, follow current aviation rules and site restrictions as well as heritage permissions. Our AI drone operations guide covers the operational control layer.
Use Artifact Vision to Triage, Not Date by Photograph
Computer vision can group pottery fabrics, match fragments, find duplicate catalogue photographs or suggest records with similar decoration. It can make a specialist queue more manageable.
An object’s archaeological value also depends on stratigraphic context, material, manufacturing evidence, associated finds, conservation state and the circumstances of discovery. A confident image label should not overwrite the original field description or assign a precise date unsupported by the assemblage.
Use controlled vocabularies and let the system abstain. Show similar reference examples and why they were retrieved. A specialist should approve the catalogue term, confidence and any chronology before publication. Track corrections back into the evaluation set, but never train on a correction without preserving who made it and under which catalogue standard.
Finds law also requires a jurisdiction gate. The Treasure Act Code of Practice, third revision applies to England, Wales and Northern Ireland and reflects the 2023 expansion of the definition; Scotland operates a separate treasure-trove system. If there is doubt, use the relevant Finds Liaison Officer, museum or national authority route. Classification software must not decide that a find is outside a reporting duty.
A 3D Model Is Not the Preserved Site
Photogrammetry and laser scanning can record shape, colour and condition at a point in time. AI can assist alignment, segmentation, damage comparison and navigation. The result is a measurement-derived representation—not a “perfect digital twin” and not a substitute for physical conservation.
Record scale, control points, equipment, calibration, capture geometry, lighting, software, processing, coordinate system, accuracy assessment and areas with poor coverage. Retain source photographs or scans where rights and storage policy allow. Distinguish measured geometry from reconstructed or generated surfaces, and make uncertainty visible to viewers.
Use repeat surveys for condition monitoring only after registration error and survey variation are quantified. A colour or surface change may come from light, moisture, camera settings or alignment rather than deterioration.
Our AI museum-curation guide covers interpretation and public display after the archaeological record has been established.
Plan the Digital Archive Before Capture
A hard drive full of point clouds is not preservation. Long-term reuse needs selected files, stable formats, metadata, rights, identifiers, checksums and a funded repository.
The Archaeology Data Service’s 2026 data-management training covers documentation and long-term preservation across the research lifecycle. Its updated depositor instructions specify dataset structure, file naming and metadata preparation. Agree the data-management and selection plan before expensive capture begins.
Archive:
- raw or authoritative source data where deposit rights permit;
- processing workflow and parameters;
- model code or sufficient version and configuration information;
- training and evaluation provenance within legal and ethical limits;
- candidate and rejection layers;
- expert decisions and field-verification results;
- final reports, GIS and catalogue outputs; and
- rights, access restrictions and preservation metadata.
Sensitive locations may require restricted access. Open data is not an excuse to expose vulnerable sites, human remains or personal data.
For broader archive design, continue with AI for libraries and digital preservation.
From Woodland Candidate to Historic Record
Imagine a model flags a circular platform in public lidar. A weak workflow publishes “new prehistoric monument” with a confidence score.
A governed team first preserves the source tile and processing. The archaeologist reviews slope, forestry history, nearby records and alternative explanations. The project checks access and designation constraints, then uses a non-invasive field survey. The feature proves to be a modern charcoal-burning platform, still historically interesting but not prehistoric.
The candidate, revised interpretation, field notes and location enter the appropriate record and archive. The model’s error becomes useful evidence for the next woodland evaluation rather than disappearing from a marketing claim.
Release Gates for an Archaeological Pilot
| Gate | Pass condition before scale |
|---|---|
| Research | Named question, feature classes, landscape, decision and prohibited uses |
| Source | 100% of inputs have provenance, licence, spatial reference and processing record |
| Evaluation | Geographic hold-out results meet class-specific thresholds; false candidates fit reviewer capacity |
| Uncertainty | “Not assessable” and alternative explanations are available; no-detection is never rendered as absence |
| Consent | Every field action has recorded access, designation and method approvals |
| Finds | Reporting route and archive owner are defined before recovery; AI cannot clear a reporting duty |
| Record | Expert interpretation, verification and corrections link back to the candidate and source |
| Preservation | Data-management plan, accepted formats, metadata, repository and costs are agreed |
| Public claim | Every announced discovery has specialist approval and a stable supporting record |
Monitor reviewer hours per confirmed feature, false candidates by landscape, field confirmation and reinterpretation rates, missing provenance, archive rejection, consent exceptions and time from verification to record deposit. Pause if review backlogs turn unassessed candidates into apparent facts.
The Durable Discovery Is a Reusable Record
AI can help archaeologists search terrain, images and catalogues at a scale that manual review alone cannot match. It should increase the number of defensible questions, not the speed at which uncertain patterns become headlines.
Preserve the source, uncertainty and expert decision. Verify proportionately and lawfully. Deposit the result where future researchers can understand how it was made. Archaeology is not just finding a feature; it is keeping enough context for that feature to remain knowledge.
