Gains:
- Being able to distinguish which steps of the geomatics workflow saves real time and which decisions should remain with the approval of the competent engineer
- Ability to implement a discipline that cross-validates each AI output with order of magnitude, geodesic plausibility, and terrain/resource evidence
- Acquire the habit of setting up secure prompts by recognizing the risks of coordinate/datum confusion, hallucination and personal data privacy and anonymizing the context.
Geomatics engineering is the branch of engineering that measures, processes, analyzes and maps the position of the earth and the objects on it. This profession works with coordinates, millimeter to centimeter accuracy, property boundaries and official documents; So mistakes are expensive. In this unit, we establish where artificial intelligence (AI in short: software systems that can process text, images and data and produce human-like answers) saves real time in the geomatics workflow, where it is dangerous and how to verify each output. Let's say it from the beginning: artificial intelligence is an assistant, although it produces the coordinates, the competent engineer makes the decision and signature.
First, let's clarify where profession and AI intersect. A geomatics project typically goes through the following steps: data collection (land surveying with GNSS, total station, UAV, LiDAR), data processing (converting raw observation to coordinates, balancing), analysis (spatial queries with GIS/GIS), production (map, orthophoto, model), delivery (report, sheet, registration). Artificial intelligence works at different levels in each link of this chain: in some places it does the main job (for example, classifying thousands of images), in some places it only produces drafts (report text), and in some places it is an assistant that should not be trusted at all (the exact value of a cadastral border).
The Principle That Determines What to Give to AI
You can measure the risk of every AI task in one sentence: the risk of an output is equal to the harm it will cause if that output is faulty. A minor error in a literature summary is almost harmless; However, a 20 cm error in the coordinates of an application (marking the points in the project on the land) means a foundation poured in the wrong place, a building overflowing onto the neighboring parcel and a dispute in court. So we scale verification intensity according to potential harm.
To make this concrete, let's break things down into three buckets:
business type
The role of AI
Verification density
example
Low risk/helpful
original manufacturer
Lightweight (review)
Converting measurement notes into tables, report draft
Medium risk/draft
accelerator
Medium (sample control + logic)
Land cover preliminary classification, code generation
High risk/critical
It just gives an idea
Heavy (full independent verification)
Cadastral border, application, elevation value, datum conversion
This table must be a reflex: when you receive an AI output, you first ask "which bucket is this in?" ', then spend your verification energy accordingly.
Verification Discipline: Three Filters
In Geomatics, pass each AI output through three filters.
- Magnitude control. Is the number within a reasonable range? The ED50/UTM easting value of a point in Türkiye is approximately 6 digit meters; degrees of longitude range from -180 to +180; The area of a parcel is in square metres. If the AI returned a 3-digit result for an expected 6-digit right value, you would know there was an error without reading the content of the answer. This is the cheapest and most powerful strainer.
- Geodetic/spatial plausibility. Is the result compatible with physics and geometry? The distance between two neighboring points cannot be negative; The sum of the interior angles of a closed polygon must converge to a certain value; The result of a buffer operation must be larger than the input geometry. If the AI said something that is logically impossible, it is eliminated.
- Independent evidence. Compare the output with a source the AI has not seen: a known control point, ground truth, an official document, a second method calculation. An example of this is testing a transformation result with a reference whose coordinates are already known.
Tip: Apply the three filters in order, but never skip the first two; because the order and plausibility check takes seconds and catches most errors before they even go to independent evidence.
Coordinate, Datum and Privacy: Pitfalls Specific to Geomatics
Besides general AI risks, this profession has two major pitfalls unique to it. The first is the coordinate/datum confusion. Datum is a mathematical reference surface representing the earth's surface (e.g. WGS84, ED50, ITRF); The same physical point is expressed with different numbers in different datums, and the difference between them can be meters in some places in our country. If you don't specify which datum you're talking about, the AI makes an assumption and silently converts it to the wrong system. Second, personal data and property privacy: cadastral records contain owner name, identification number, precise location; Sending them uncontrolled to an external model is against both legislation and professional ethics.
Caution: Stop before pasting the real owner's name, TR ID number, or the exact coordinates of a sensitive facility into a prompt. Anonymize (make names “Malik-1”, parcels “Island/Plot-A”) or use an in-house/approved environment.
Three Mini Cases: By the Numbers
Case 1 — The rank filter saves lives. An intern had the AI project a point from a geographic coordinate and got a result of 4,281 m to the right. In Türkiye, UTM right values are in the 100,000-900,000 m band; A 4-digit result was impossible. With just the order check, the error was caught immediately because the model assumed the output was radians rather than degrees; The coordinate was corrected before transferring it to CAD.
Case 2 — Datum drift. A team took 12 control points from an old map (ED50) and gave it to the AI as "convert to UTM", without specifying the datum. Assumed model WGS84; The results systematically shifted approximately 3-4 m from GPS measurements in the field. When compared to a single known reference, the systematic difference was immediately apparent and the job was repeated with the correct transformation parameters. One-point independent control saved the entire campaign.
Case 3 — Preventing privacy breach. One office would load a cadastral database of 850 parcels into an external model to "find inconsistencies"; The table contained owner names and identification numbers. Thanks to the checklist, they first anonymized it: personal columns were removed, only geometry and area information were sent. The benefit of the analysis remained the same, personal data never got out.
Weak Prompt / Strong Prompt
Weak prompt:
Convert these coordinates to UTM and write a report.[coordinates]
Powerful prompt:
Your role: geomatics data assistant. Task: convert the following points from the SOURCE system to the TARGET system.- Source CRS: EPSG:4326 (WGS84 geo, degree)- Target CRS: EPSG:5256 (TUREF / TM33)- If there is unclear or missing information, DO NOT CONVERT, ask first.- Specify the expected right value rank (6 digits) for each point.- Note that the output should be validated, compared with a known control point.Data (anonymous): [dot table]
The powerful prompt clearly states the source and target, commands to stop at incomplete information, and rewrites the rank expectation and verification obligation. Thus, the output is both more accurate and auditable.
Four Copiable Templates
1) Job classification filter:
Place the following task into one of three buckets: (a) low risk/helpful,(b) medium risk/drafty, (c) high risk/critical. Write your rationale and suggested verification steps. Task: [job description]
2) Rank control request:
Write the expected order of magnitude (unit and typical range) before producing the result below. If the result is outside this range, issue a "RANGE WARNING" and review the calculation. Input: [data]
3) Pre-anonymization scanning:
Does the following text contain personal data (name, ID number, title deed owner information, sensitive location) or confidential information? If so, list them and suggest anonymous equivalents. DO NOT change the text, just report it. Text: [text]
4) Verification plan generator:
Write a verification scheme for the following AI output: (1) rank check, (2) geodesic/spatial plausibility, (3) independent source of evidence. Write clearly what will be taken into account at every step. Output: [AI output]
Common mistakes
- Operating coordinates without specifying Datum/CRS. The most common and most expensive mistake; It causes silent slippage for meters.
- Mistaking fluency for accuracy. AI writes confidently and neatly; This does not indicate that the content is accurate.
- Bypassing the rank check. Bringing errors that a split-second glance would catch into CAD.
- Sending personal/property data without control. The justification "useful anyway" does not remove the obligation of confidentiality.
- Basing critical work on a single AI output. Proceeding with works such as application, cadastral border and elevation without independent verification.
- Thinking that you are transferring responsibility to the model. The signature and responsibility always remain with the competent engineer.
In summary
Artificial intelligence works with different weights at each step of the geomatics workflow; The critical thing is to consider the risk of each mission: "What do I lose if there is an error?" question and scale the verification accordingly. Three filters (degree, plausibility, independent evidence) catch most errors cheaply. Two major pitfalls specific to this profession are coordinate/datum confusion and personal/property data confidentiality; Both are managed by establishing a disciplined prompt from the beginning. The final decision and signature remain with the competent engineer.
Application task
Choose three geomatics tasks from your own business (or hypothetical): one low risk, one medium risk, one high risk. For each, write down (1) which bucket it falls into, (2) the role of the AI, (3) the three filter steps you will apply. For the high-risk task, specify concretely what independent evidence (checkpoint, document, second method) you will test the output against.
checklist
- [ ] I measured the risk of the mission by "damage that will occur if there is an error."
- [ ] I placed the task in the low/medium/high risk bucket.
- [ ] I have explicitly specified the source and target coordinate system (datum/EPSG) in the prompt.
- [ ] I passed the result through the rank check.
- [ ] I tested geodetic/spatial plausibility.
- [ ] I have verified it with at least one independent evidence.
- [ ] I have anonymized personal/proprietary data or worked in an approved environment.
- [ ] I made the final decision and signature the responsibility of the engineer.