Gains:
- Ability to scan for anomalies in satellite, aerial photography and LiDAR data with artificial intelligence and evaluate each sign as a candidate requiring terrain confirmation
- Understanding the seasonal and conditional dependence of plant, soil and shadow traces and how natural/modern false positives are eliminated.
- Ability to apply why sensitive site locations should be protected against plunder and blurred in broadcast
One of the most exciting developments in archeology is the ability to see traces underground and on the surface without digging in the ground. Remote sensing (a method of examining an area with satellite, aircraft or drone images without physically touching it) is the most powerful tool for scanning large areas and detecting possible sites (places containing archaeological remains). In this unit, we will learn how AI marks anomalies (tracks that differ from its surroundings, which may be man-made) in satellite images, aerial photographs and LiDAR data, but why each mark must be confirmed in the field.
Basic principle: AI is a scanning and pointing assistant that says "look here" in a wide area; "This is really a site" decision is made by the land control and the expert. In remote sensing, every point marked by AI is a hypothesis, not a discovery.
Types of remote sensing
Satellite and aerial photography. Buried structures lead to differential growth of the soil and vegetation above them. The crop on a wall is weaker, on a ditch it is bushier; These are called crop marks. NDVI (an index that measures the health of vegetation; it quantifies plant vitality using infrared light) highlights these differences. AI scans these subtle differences in large images faster than a human.
LiDAR. LiDAR (Light Detection and Ranging — technology that creates a very precise 3D elevation map of the surface by shooting a laser from an aircraft to the ground and measuring the return time) is the most revolutionary tool. Its feature is that the laser can reach the ground through trees; Thus mounds, terraces, roads and building foundations under the forest are revealed when the vegetation is numerically "peeled". AI marks bumps, pits and smooth geometric shapes in this elevation data.
Geophysics (supplementary). Methods such as ground penetrating radar (GPR) and magnetic surveying see just below the surface; AI helps look for anomaly patterns in this data as well.
Site detection steps with AI
- Data preparation. Images are fitted to the coordinate system (georeferencing), shadow and contrast are corrected. In LiDAR, vegetation is filtered out and a "bare ground model" is produced.
- Anomaly scanning. AI marks smooth geometries that may be man-made (circular mound, rectangular structure, linear ditch/road).
- Elimination. Most of the signs are natural (geological, agricultural, modern). The expert eliminates those that are clearly natural and modern.
- Prioritization. The remaining candidate points are prioritized by context and known site distribution.
- Land confirmation. Priority points are checked on site by surface survey and, if necessary, drilling. No point is declared "sit" without confirmation.
Tip: Always see signs of AI as “a list of candidates with a high false positive rate.” The aim is to intelligently narrow down the places to visit in the field; not to announce the discovery. Reducing 200 land points to 30 in a season is a big gain.
Sensitive location: critical privacy
The most serious ethical and security issue in site detection is location confidentiality. The precise coordinates of a newly identified, unprotected site are a direct target for illegal diggers and treasure seekers. Therefore:
- Do not enter public AI vehicles whose exact coordinates have not been confirmed.
- Blur the location of sensitive sites in broadcasts and presentations (give general area, no GPS).
- Keep data in the organization's secure system and limit access.
Attention: Pasting the coordinates of a site into a public service saying "I'm just going to have a map drawn" may leave that site open to plunder. Location data is an asset that must be protected as much as the find itself.
three mini cases
Case 1 — The scan narrowed the area. A survey project would survey an area of 500 km². AI-powered LiDAR analysis marked 120 possible mounds under the forest. After expert elimination, 38 priority spots remain; 11 of these were found to be real archaeological structures in the field. AI has condensed months of blind searching into a targeted season.
Case 2 — False positive was the lesson. While a team was excitedly preparing to declare a smooth circular scar marked by the AI a "mound", the expert geologist showed that it was a natural depression (dolin). A statement made without land confirmation could be a loss of reputation. Lesson: AI sees geometry, not interpretation of origin.
Case 3 — Location leak prevented. A student would upload the image containing the coordinates of a newly found site to a public tool. The counselor stopped; the area was still unprotected. The image was processed with coordinate information removed and only in the secure system.
Four copyable templates
1) Anomaly scanning frame:
Your role: remote sensing analysis assistant. I will give you a bare-ground (LiDAR derived) elevation image/data. Mark smooth geometries (circular, rectangular, linear) that MAY be man-made; For each, write down its location, format, and why it is notable. For each sign, also indicate the possibility of a natural/modern explanation. Don't declare a definitive "sit".
2) False positive elimination:
Below is the list of candidate points that the AI has marked. For each, list the natural and modern explanations (geological formation, agricultural trace, road, building foundation) that I can eliminate BEFORE the terrain check. List the remaining "really vague" candidates separately.
3) Land control plan:
My top candidate spots are: [list]. For each point, a checklist appears containing the indicators I should look for when confirming in the field (surface findings, topography, soil color, stone density) and safe/ethical field rules (permission, location confidentiality).
4) Location blurring for broadcast:
When preparing the following site information for a publication/presentation, suggest how I might describe it WITHOUT revealing the exact location: list what level of generalization (province/region), what information should be omitted. The goal: a balance of scientific transparency and protection against spoilage.
Weak prompt / Strong prompt
Weak prompt:
Is there an archaeological site in this satellite image? Give its coordinates.
AI can declare a natural trace a “sit” and fit a confident coordinate; This is dangerous from both a scientific and a security perspective.
Powerful prompt:
Your role: remote sensing assistant. Mark the smooth geometric marks that MAY be man-made in this image. For each trace: (a) its shape, (b) why it is remarkable, (c) natural/modern alternative explanation. These are candidates to be checked ON THE FIELD only; do not declare a "sit". Generate exact coordinates, describe relative position.
Difference: the first makes discovery declared; the latter produces a secure list of candidates to be verified.
Image types and seasonal effects
In remote sensing an image does not "always" work; It takes the right condition to see the right track. Crop marks appear best during dry periods when the crop is under water stress: the soil above a buried wall retains less water, so the crop there turns yellow early and appears as a line from above. In a humid spring image, the same trace may not appear at all. Soil marks appear when different colored soil comes to the surface after the field is plowed. Shadow marks, on the other hand, reveal light bumps and potholes with long shadows in low sun, in the morning or evening. Therefore, it is necessary to evaluate an area not with a single image, but with many images in different seasons and times; AI accelerates this multi-time comparison.
The advantage of LiDAR is that it largely overcomes this condition dependence: since it directly measures surface topography, it is less dependent on the season of vegetation and can see the forest understory. However, when producing a "bare ground model" in LiDAR, if the plant filter is too aggressive, it can erase the real bumps, and if it is too soft, it can mistake the trees for structures; These filter decisions require expert supervision.
Tip: When evaluating an area, ask "in which season, at what time, and with what method?" Always ask the question. If there is no trace in a single image, it does not mean that there is no site; Maybe you're looking at it in the wrong condition.
Method comparison table
Method
what sees
Contribution of AI
border
Satellite/aerial photography
Plant/soil traces
Wide area scanning
season, resolution
NDVI
Plant vigor difference
Scar highlighting
Confuses natural differences
LiDAR
surface topography
forest understory geometry
Plant filter errors
Geophysics (GPR/magnetic)
subsurface structure
Anomaly pattern
local, slow
Common mistakes
- Mistaking an anomaly for a discovery. The sign is a candidate; Without land confirmation it is not a site.
- Not eliminating natural/modern traces. Dolins, agricultural scars and pipes frequently appear as "structures".
- Giving the exact coordinate to the open vehicle. Risk of looting; Location is an asset that must be protected.
- Relying on a single image. Interpretation is incomplete without different seasons, different methods and terrain.
- Forgetting the false positive rate. Most signs come up empty; This is normal, the purpose is narrowing.
In summary
In remote sensing, AI saves tremendous time in scanning large areas and marking possible sites. But every sign is a hypothesis that must be confirmed in the field; natural and modern traces are carefully eliminated; and most importantly, sensitive locations are protected from looting. AI says “look here”; The land and the expert decide whether "this place is protected".
Application task
Choose an open-access LiDAR or satellite image (from your own region or from a sample dataset). Have the AI flag possible geometries with the “Anomaly scanning” template, then eliminate natural/modern explanations with the “False positive elimination” template. Draft a land control plan for the remaining candidates and include a location confidentiality note.
checklist
- [ ] I considered the AI signals to be "candidates", not discoveries.
- [ ] I tried to eliminate natural and modern traces.
- [ ] I have planned land confirmation for priority candidates.
- [ ] I left the sensitive locations out of the open vehicle and blurred them in the broadcast.
- [ ] I took into account that the false positive rate may be high.