Gains:
- Ability to explain the contribution of artificial intelligence in the process of producing scaled 3D models, point clouds and orthophotos with photogrammetry and why the scale bar is essential.
- Recognizing the risk of fake surfaces created by filling in missing data and being able to visually separate evidence from completing the prediction.
- Understanding that a model is a measurement, and interpretation is context and expert judgment, such as what is a wall and what is natural stone.
In archaeology, excavation is a destructive act: when you excavate a layer, you permanently destroy it and you will never see it the same way again. Therefore, documentation (making a permanent record of each stage of the excavation with measurements, drawings, photographs and models) is one of the most sacred duties of archeology. In this unit, we will see how photogrammetry (a technique that produces a scaled three-dimensional model by matching common points from multiple photographs of an object taken from different angles) and how AI accelerates 3D documentation, but why decisions on scale and interpretation remain with humans.
Basic principle: AI helps produce a scale model, orthophotos and measurements; But what to document, which line is a wall and which is natural stone, and how to interpret the model are expert decisions. The 3D model is a proof; its reading is an interpretation.
Steps of photogrammetry and 3D documentation
1. Photo shoot. The object or area is captured with sufficient overlap (each point in at least three photographs) and in good light. Scale bar and known distances (control points) are placed next to it; Without scale the model is dimensionless.
2. Alignment and point cloud. The software matches common points in the photos and produces a point cloud (data consisting of millions of 3D points representing the object's surface). AI speeds up this matching and noise cleaning.
3. Mesh and texture. The points are converted into a surface (mesh) and a photo texture is applied on top; realistic 3D model emerges.
4. Orthophoto and measurement. A scaled orthophoto (a top view with every point at the correct scale) with perspective distortion removed is extracted from the model. Survey drawings are made on this.
5. Surveying and interpretation. Surveying (a scaled, interpreted technical drawing of a building or ruin) is the stage where the expert reads the model and distinguishes what is a wall, what is mortar, and what is a later addition. AI can suggest edges, but the expert makes the distinction.
Tip: Never skip the scale bar and control points. AI produces a "beautiful" model from unscaled photographs, but that model is scientifically worthless because real measurements cannot be taken from it. The value of documentation is in its measurability.
Contribution and limits of AI
AI; It is powerful in photo alignment, noise removal, missing surface filling (interpolation) and comparison of large numbers of models. But there are two critical limits:
- Risk of fabricated surface. AI can fill in missing data in a way that looks “plausible”; that is, it may produce a surface that was not actually observed in the model. This is dangerous in documentation: a detail that does not actually exist appears as evidence. Filled/predicted regions should always be marked.
- Measurement, not comment. The model does not say what is what. Whether a pit is a grave or a garbage pit, whether a row of stones is a wall or a road, comes from the context and expert judgment, not from the model.
Caution: Every piece the AI adds when “restoring” or “completing” a 3D model is an assumption. If the evidence (the actual measured surface) and the reconstruction (the estimated completion) are not visually separated, the viewer will mistake the assumption for the truth. This is documentation's most serious ethical error.
three mini cases
Case 1 — Fast documentation saved. On a rescue excavation (an emergency excavation due to construction), the team had only a few hours to document a layer. Drone and AI-accelerated photogrammetry produced a millimetric 3D record of the layer in 45 minutes; The layer was removed the next day but the pattern remained permanent. Hand drawing was impossible during this time.
Case 2 — Fabricated surface caught. A student would write in the report that the flat surface filled by the AI in the model he created with a few photographs was a "solid base". The consultant showed that there were not enough photographs in that area and that the surface was an interpolation. The model was reorganized by separating "measured" and "predicted" regions.
Case 3 — The scale-free model did not work. One team delivered a building model that looked nice but was shot without a scale bar; No real measurements could be taken from it and no survey drawings could be made. The shooting was repeated with checkpoints. Lesson: scale is the scientific value of the model.
Four copyable templates
1) Shooting plan control:
Your role: photogrammetry documentation assistant. I will document the following object/area: [description]. Give me a shooting checklist for aliasing, number of angles, lighting, scale bar and control point placement. Highlight critical points for scalability and full coverage.
2) Model quality inspection:
Below I will give my 3D model production parameters/statistics (number of photos, overlap, area covered). In what areas of the model might the data be weak and likely to be filled (predicted) by AI? Tell me how I should mark these.
3) Evidence/fiction distinction:
I'm preparing a 3D reconstruction/completion. Suggest a marking scheme (color/layer/label) to distinguish which parts of the model are MEASURED evidence and which are ESTIMATED completion. Also draft a statement text that will prevent the viewer from mistaking the assumption for fact.
4) Preparation for survey:
I will draw a survey from the orthophoto. List the elements I need to distinguish in this structure (wall/mortar/addition/repair/natural stone) and the visual clues to pay attention to for each. I will make the definitive comment; You give the criteria to be checked.
Weak prompt / Strong prompt
Weak prompt:
Create a 3D model of the building from these photos and fill in the missing parts.
“Fill in missing parts” allows the AI to generate a fitting surface; The result is a misleading model in which evidence and fiction are confused.
Powerful prompt:
When creating a scale 3D model from these photographs, ONLY adequately model the surfaces covered by the photograph. FILLING data poor areas; mark them as "missing data". Report how many photos support each region in model statistics. Verify dimensions against scale bar.
The difference: the first produces a beautiful but false whole; the latter produces an honest document with visible gaps.
Why won't 3D documentation fully replace traditional drawing?
The 3D model is powerful but does not completely invalidate traditional hand drawing and observation notebook; the two complement each other. A survey or layer drawing contains the expert's interpretation at the time of excavation: which stone belongs to the wall, which was rolled later, whether a line is a real boundary or a shadow. This interpretation happens in the illustrator's head and does not appear automatically in the 3D model. The model measures “whatever it is”; But the distinction between "which of these makes sense" is human observation.
Additionally, the 3D model carries an archive format problem: the model file you produce today may not be open ten years from now; software and format become obsolete. That's why raw photographs, control point measurements and a standard drawing/orthophoto are always stored together. The model is not the only leg of long-term documentation, but the richest leg.
Another practical issue is coordinates and orientation: a model is just a shape "in the air" if it is not grounded in a real-world coordinate system (georeference); He loses which point of the excavation he is looking at and in which direction. Measuring control points with actual coordinates links the model to the excavation plan and enables models from different seasons to be superimposed.
Tip: Don't just generate the 3D model and leave it at that; Include raw photos, control point coordinates, scale information, and an annotated drawing. The future researcher should be able to understand your work with these additional documents, even if they cannot open the model.
3D documentation method table
Stage
Contribution of AI
human decision
verification
shooting
Plan/coverage recommendation
scale, checkpoint
Aliasing control
point cloud
Pairing, noise removal
Quality threshold
Control point size
surface/texture
automatic mesh
Filling limit
Weak zone sign
Orthophoto/measurement
scale production
Scale verification
known distance
survey
Edge suggestion
Wall/mortar/attachment separation
context, expert
Common mistakes
- Skip the scale bar. A model without scale is not scientifically measurable.
- Filling in missing data. AI produces fitting surface; Gaps must be marked.
- Not separating evidence from fiction. It is misleading if the completed regions are not visually separated.
- Mistaking the model for an interpretation. The model meter does not say what it is; Interpretation is a matter of context and expert work.
- Creating a model with few photos. Insufficient overlap gives a distorted and gapped model.
In summary
AI in 3D documentation and photogrammetry; It produces fast, millimetric and repeatable recording and is especially invaluable in time-pressured rescue excavations. But the scale bar is essential, missing data is not filled in but marked, evidence is separated from fiction, and the interpretation of the model (what is what) belongs to the expert. The model is a proof; Reading is a responsibility.
Application task
Take a small find or corner of a building with the scale bar, at least 20 overlapping photos. Generate a model with a photogrammetry flow and identify weak/filled areas with the "Model quality check" template. Then create a diagram with the “Evidence/fiction distinction” template, marking which regions are measured evidence and which are guesses.
checklist
- [ ] I used scale bar and control points in the shot.
- [ ] I took it with adequate overlap (3+ photos per spot).
- [ ] I marked the weak/filled areas, I did not leave a fake surface.
- [ ] I visually separated the proof from the completion of the prediction.
- [ ] I interpreted the model using context and expert judgment.