Gains:
- Understanding the logic of AI-supported detection of urban sprawl, construction and land cover change through remote sensing and satellite imagery.
- Ability to critically evaluate image classification and change detection outputs with accuracy, resolution and date information
- Ability to understand that the boundaries that AI extracts from the image must be confirmed by field and official records and cannot be legal evidence.
It is impossible to go to the field one by one to see where, how much and how fast a city has grown in the last decade. Instead, planners use remote sensing (imaging the Earth using satellites and aircraft and extracting data). By comparing satellite images from different dates, you can detect new construction, lost agricultural land, expanding illegal construction or shrinking green space. AI is a powerful accelerator in classifying these images (labeling each pixel as “structure,” “plant,” “water,” etc.) and detecting changes between two dates. But image analysis is full of resolution, date, shadow, cloud and classification error. A “change” that the AI flags could be a shadow, a seasonal difference, or an error. Therefore, no conclusion drawn from the image can be used as a basis for legal action without being confirmed by field and official records.
Basic concepts of remote sensing
Resolution: The area occupied by a pixel on the ground. At 10 meter resolution, one pixel represents 10x10 meters; A small outbuilding is not visible. High resolution (e.g. 0.3-1 meter) means more detail but more expensive and less frequent data.
Spectral band: Satellites also record light that the eye cannot see (such as infrared). Vegetation glows in the infrared; this makes it easier to separate the plant from the structure. Indices such as NDVI (a simple calculator that measures plant vigor) are common in greenfield tracking.
Date and season: If two images are in different seasons, the decrease in green may not be a "loss" but the arrival of winter. Date and season harmony are essential in comparison.
Accuracy: The value that measures how accurate the classification is. “85 percent accuracy” means that one in every six pixels may be mislabeled; This is a significant margin of error for an audit process.
Attention: A structure detected by AI from a satellite image is not, by itself, evidence of an illegal structure. Shadow, resolution, date and classification error produce false positives. Inspection and enforcement always require site control, title deed/license registration and official measurement.
The biggest value that free and open satellite data (sources such as Europe's Sentinel and the US's Landsat programs) bring to planning is the historical time series: you have access to dozens of images of the same area spanning years. This allows you to go beyond a single "before/after" comparison and see a neighborhood's sprawl rate, seasonal fluctuation, and trend. AI is helpful in summarizing these long series and separating meaningful patterns of change from noise. But the same rule applies in time series: no matter how clear a trend appears, it is the resolution of the images, cloud rate and seasonality that determines the outcome; No trend can be considered reliable without reporting these three factors.
Logic of change detection
There are several ways to compare two dated images: classifying both dates and comparing classes, getting index difference (e.g. NDVI difference), or directly producing a "changed/not changed" map with the AI model. The result is a map of change: where is the new structure, where is the plant lost. This map is invaluable as a preliminary screening — it tells the planner to "look over there." But each sign is separately verified whether it is a real change or an error.
Step by step: AI-powered change analysis
- Define the question. “Where is the new construction in [the area] between 2018 and 2024?”
- Select image. Similar season, acceptable resolution, low cloud.
- Preprocessing. Bring it to the same CRS, align, apply cloud/shadow mask.
- Classify/compare. Generate class or change map with AI.
- Evaluate accuracy. Measure accuracy with known reference points.
- Field and registration confirmation. Check marked changes with title deed/license and site if necessary.
- Report limits. Write clearly the date, resolution, accuracy and margin of error.
Weak prompt / Strong prompt
Weak prompt: "Tell me what changed from these two images."
AI produces a nice-looking but uncertainly accurate list; You may think the difference in shade and season is real change.
Strong prompt: "I am comparing classified land cover data from two dates (July 2018 / July 2024, same season, 10 m resolution). When reporting change analysis, state clearly: which changes may be false positives due to shade/season/resolution, which determinations require field verification, and how accuracy should be measured. Do not be definitive; note each finding with 'field confirmation required'."
The second prompt forces the AI to clearly communicate uncertainty and sources of error.
Four copyable templates
Task: Critically evaluate the following change detection output. List possible ERROR sources for each detection: shadow, cloud, season, resolution, classification error. Add note "Field confirmation required". Do not make final judgments. Findings / date / resolution: [...]
Task: Write a checklist of image selection criteria for an urban change analysis: seasonality, resolution, cloud rate, date range, CRS. Create questions to ask whether the images I have meet these criteria. Image information: [...]
Task: DRAFT a report from the following change analysis findings. Clearly state resolution, date, accuracy, and margin of error in each finding. Frame the results as “pre-screening”; Emphasize that field and official record confirmation is required for inspection/sanction.Findings: [...]
Task: Suggest a simple way to evaluate classification accuracy: how many reference points, how to choose, how to calculate accuracy (plain explanation). Which truth statement should be written in the report, example.Context: [...]
A table: data type and usage
Data
resolution
strong point
limit
medium resolution satellite
10-30m
Free, frequent, wide area
Misses small structure
high resolution satellite
0.3-1m
Detailed structure detection
Expensive, sparse, cloud risk
Aerial photo / orthophoto
0.1-0.3m
Very detailed, formal
Update frequency is low
UAV (drone)
cm level
Instant, very detailed
Small space, permit/legislation
three mini cases
Case 1 — Shadow illusion. A municipality sees "new construction" signs in a neighborhood on the AI change map. The field team goes; Most of the signs are long building shadows in the afternoon image. The actual number of new builds is half the signs. AI had pre-screened, but if used without confirmation, it would produce false minutes.
Case 2 — Escaping small structure. A planner tries to track small warehouses in an agricultural field with a free satellite with 10 meter resolution. Since the warehouses are 6x8 meters, they remain under one pixel and are not visible. The planner switches to high-resolution orthophoto for actual tracking. Resolution determines what can be seen.
Case 3 — Transparent report. One team analyzes the 6-year spread of a region and clearly writes in the report that "detections are a preliminary scan with 82 percent accuracy; 47 points await field confirmation; image dates and resolution are attached." The parliament evaluates the result with the correct weight. Reporting uncertainty rather than hiding it makes analysis reliable.
Common mistakes
- Mistaking detection as evidence. The image is pre-scanned; It requires sanctions and official registration.
- Season/date mismatch. Different seasonal images produce false variation.
- Ignoring resolution. Small structures are not visible at low resolution; It is not "none".
- Failure to report accuracy. Hiding the margin of error leads to wrong decisions.
- Forgetting shadow and cloud. The two most common sources of false positives.
- CRS mismatch. Unaligned images change in the wrong place.
In summary
Satellite and aerial imagery are the most powerful ways to see urban change at scale and speed; AI accelerates classification and change detection. But resolution determines what can be seen, season and date determine what is real, shadow and cloud determine what is a false positive. Accuracy should always be measured and reported. Most importantly: the detection from the image is a pre-scan; Field control, title deed/license registration and official measurement are essential for inspection and sanction. AI speeds up vision, not decision making.
Application task
Consider a pair of images (or scenarios) with two different dates. (1) Check the image selection criteria with the second template: season, resolution, does the cloud fit? (2) Assume a list of change detections and extract possible error sources for each detection with the initial template. (3) Produce a draft report with the third template and add accuracy and margin of error. (4) Write down which determinations will require field confirmation and how you will do this.
checklist
- [ ] I checked the season, date and resolution compatibility of the images.
- [ ] I aligned the images to the same CRS; I took the cloud/shadow mask into consideration.
- [ ] I treated change detections as preliminary screening, not as evidence.
- [ ] I evaluated and reported classification accuracy.
- I confirmed the changes marked [ ] with field and official records.
- [ ] I noted what the resolution limit might miss.
- [ ] I clearly wrote the date, resolution and margin of error in the report.