Did you know? Satellites can observe the same region of Earth repeatedly. When analysts compare images from different dates, they may notice new buildings, changes in vegetation or the spread of floodwater. AI can help turn those enormous image collections into candidates for closer inspection.
This is called satellite-image change detection. It combines Earth observation, careful comparisons across time and, in some workflows, machine-learning models. The surprising part is that a changed pixel is not necessarily a changed building, forest or coastline.
Quick answer: choose the place and dates → obtain suitable imagery → correct or filter incompatible observations → align the images → estimate the differences → classify or prioritise changes → verify the results. Computer vision can help, but a useful map needs accurate source data.
The 30-second guide
- Satellite imagery: measurements gathered by sensors looking at Earth from orbit.
- Spectral band: a measurement of energy within a particular wavelength range, such as red or near-infrared.
- Spatial resolution: the ground area represented by a pixel under a specified sensor product.
- Cloud mask: an estimate of locations obscured by clouds or related interference.
- Change detection: a process for locating meaningful differences between observations.
- Ground truth: suitable independent evidence about what actually existed at a place and time.
Not every image analysis technique uses AI. Simple image subtraction and spectral-index calculations are established methods. Machine learning can help recognise more complex patterns, classify changes and filter false alerts.
Step 1: Choose the dates and the question
Imagine a fictional coastal town wants to know whether a patch of land was cleared between March and September. The team needs images covering the same site at suitable times, but it also needs to define what “cleared” means.
A new roof, seasonal vegetation and exposed soil can all change how the surface looks. The intended classification must distinguish the actual question—perhaps loss of tree cover—from ordinary seasonal differences.
Two dates are a starting point, not automatically an adequate time series. If the land varies strongly with season, comparing similar seasons across different years may be more informative than comparing an unusually wet month with a dry one.
Step 2: Select a suitable satellite dataset
Earth-observation satellites measure different wavelengths and have different resolutions and revisit schedules. The Sentinel-2 catalog in Google Earth Engine describes a multispectral land-monitoring mission with visible, near-infrared and other bands at different spatial resolutions.
For example, some Sentinel-2 optical bands are sampled at about ten metres, while other bands use coarser pixel sizes. A ten-metre pixel represents a patch of ground, not a photograph that identifies the exact condition of every leaf or small object inside it.
Dataset selection depends on the task. A broad vegetation change may be visible with medium-resolution multispectral measurements, while inspecting a tiny structure could require different imagery and permissions.
Step 3: Understand why more than three colours matter
A normal colour photograph typically combines red, green and blue channels. Many satellite sensors also collect information in near-infrared and shortwave-infrared bands. Those measurements can help distinguish vegetation, soil, water and other surface properties.
Vegetation often reflects near-infrared light differently from red light, which is absorbed for photosynthesis. Analysts can use those differences in a spectral index to highlight broad vegetation patterns.
A colourful false-colour satellite image can be useful, but its colours may represent invisible wavelengths rather than the exact colours a human observer would have seen from the ground.
Step 4: Calculate a simple vegetation index
One familiar index is the Normalised Difference Vegetation Index (NDVI). Its standard formula is (NIR − red) ÷ (NIR + red), using near-infrared and red reflectance values.
Here is a fictional numerical example. Before a change, suppose NIR reflectance is 0.60 and red reflectance is 0.20. Then NDVI is (0.60 − 0.20) ÷ (0.60 + 0.20) = 0.50.
Afterwards, imagine NIR is 0.30 and red remains 0.20. The result is (0.30 − 0.20) ÷ (0.30 + 0.20) = 0.20. The NDVI has fallen by 0.30 in this example.
That decline could merit investigation, but it does not, by itself, prove deforestation. Season, water conditions, clouds, shadows, sensor corrections or crop growth cycles can also affect spectral indices. The Earth Engine NDVI tutorial explains the computation and its limitations.
Step 5: Remove clouds and shadows before comparing
Clouds can hide land completely in an optical image. Their shadows may make vegetation appear dark or changed even when the ground remained the same. Comparing a cloudy scene with a clear scene can produce an impressive-looking but false change map.
Developers use cloud-probability products, masks and suitable compositing procedures to reduce these problems. But every mask can make mistakes, and the final quality depends on the place, sensor and dates.
Google Earth Engine’s Sentinel-2 cloud and shadow masking tutorial illustrates why both clouds and their projected shadows need attention.
A useful question: Did Earth really change, or did the conditions under which we observed it change?
Step 6: Make sure the same ground location is compared
A pixel at one coordinate in an image must correspond to the same ground location in the comparison image. Small registration shifts can make a road edge or building outline appear as a narrow strip of change, even if nothing moved.
A robust workflow uses appropriately georeferenced data, consistent projection and pixel alignment, or tests the effect of any remaining uncertainty. Mixing data with very different spatial resolutions can also produce misleading apparent differences.
This is one reason a visual side-by-side comparison is not the same as a verified numerical analysis. Geometric accuracy matters before a statistical model is asked to interpret a changed pixel.
Step 7: Locate the changed regions
One basic approach subtracts corresponding spectral measurements or indices from two dates and flags sufficiently large differences. More advanced statistical methods compare information across multiple bands.
Machine-learning systems can add classification: instead of merely saying “different,” they might predict categories such as vegetation loss, new construction or increased water extent under a properly defined task.
The Google Earth Engine community change-detection tutorial demonstrates an iterative multivariate method applied to Sentinel-2 scenes and subsequent clustering. That is an example of combining mathematical change detection with categorisation.
A method’s predicted label is still an estimate. Developers need appropriate reference material and validation to know which apparent changes represent the real-world categories they care about.
Step 8: Where does AI actually help?
A supervised image model can learn from labelled examples of before-and-after patches. It may predict whether a building appeared, vegetation disappeared or a flooded region expanded. Some workflows use image segmentation to produce a spatial map rather than one label for the whole photograph.
Other techniques cluster changes without predefined labels, or use models trained to compare image pairs. Regardless of architecture, inputs must be representative and the labels must reflect genuine observations.
A network that sees only one region’s building styles may struggle elsewhere. A model trained on low-cloud scenes may fail during a storm. Testing across geography, seasons and sensor conditions is essential.
Step 9: Why several dates are often better than two
A single dark image can reflect an unusual cloud shadow or a genuinely changed surface. An appropriate time series offers more context about what happened before and after the suspect date.
If NDVI drops sharply during a normal harvest and rises again, the pattern differs from a lasting conversion to paved ground. Yet even that observation requires correct land-use context; not all crop or forest patterns repeat identically.
A reliable system should communicate whether it found a short-term appearance change or enough evidence for a longer-term land-cover conclusion.
Step 10: Measure area without pretending every pixel is certain
Suppose a fictional change mask contains 100 detected pixels and each corresponds to a ten-by-ten-metre ground square. The nominal flagged area is 100 × 100 = 10,000 square metres, or one hectare.
That arithmetic is easy, but the interpretation is not. Some flagged pixels may be false positives, and pixels can contain mixtures of land cover. The actual area estimate depends on resolution, mapping conventions and the accuracy of the change mask.
For a serious report, a team should evaluate error rates and, where appropriate, uncertainty in the area estimate rather than publishing a perfectly precise-looking number without qualifications.
What can satellite change detection help people study?
- Vegetation: identify areas that may have lost or gained visible green cover.
- Land development: examine changes in the built environment across appropriate time intervals.
- Water and flooding: flag changes in observed surface-water patterns using suitable sensors.
- Agriculture: monitor seasonal crop-related spectral patterns.
- Disaster research: prioritise candidate regions for further assessment after an event.
These are applications of analysis, not claims that a satellite image alone proves the cause of a change. Field observations and suitable independent data remain important.
Five common false alarms
- Cloud shadow: a patch of land looks darker in only one image.
- Seasonal growth: plants change colour or density without permanent land conversion.
- Different sensor processing: values shift because datasets were calibrated or harmonised differently.
- Misregistration: a road edge moves by a pixel due to alignment error.
- Mixed pixels: one pixel includes water, soil and vegetation, complicating its label.
Try it: a paper change-detection activity
Draw two four-by-four grids representing the same fictional park in June and September. Shade green cells for vegetation and grey cells for paths. Change only three cells between the grids and ask a friend to find those changes.
Now shift the second grid one square to the right before comparing. Many unchanged features suddenly appear different. The exercise shows why matching positions matters before making a change map.
Finally label one square “cloud-covered.” Can you honestly decide whether it changed? The correct answer is often “not from this observation alone.”
How should satellite findings be communicated responsibly?
Broad-area satellite analysis can support environmental and geographic research, but claims about ownership, illegal activity or the actions of particular individuals require independent evidence and due process.
A prediction map may contain uncertainty and errors. Publishing precise allegations based only on a machine-generated land-cover mask can mislead readers or unfairly affect communities.
Report source dates, bands, resolution, processing methods, ground verification and limitations. Be especially careful when images are incomplete or conditions differ across the comparison period.
Frequently asked questions
Is satellite change detection always artificial intelligence?
No. Image differencing and spectral indices can be computed without machine learning. AI can help categorise complex patterns.
What is NDVI?
A vegetation-related spectral index computed from near-infrared and red measurements: (NIR − red) ÷ (NIR + red).
Can a satellite see individual people in these datasets?
Resolution varies. Common medium-resolution land-observation products are not appropriate for identifying individuals, and such identity claims should not be invented.
Does a drop in NDVI prove trees were cut down?
No. It is a clue that needs seasonal, cloud, sensor and independent land-cover context.
Why can’t I just subtract two photographs?
You can calculate a difference, but registration, lighting, shadows, atmospheric conditions and sensor differences may create false signals.
Your one-minute takeaway
- Satellite change detection compares appropriately aligned observations through time.
- Clouds, seasons, band differences and pixel resolution can create misleading apparent changes.
- AI can help classify candidate changes, but consequential conclusions require independent validation.
Further reading and connected AI guides
- Google Earth Engine: Sentinel-2 Datasets — band information, resolution and harmonised data.
- Google Earth Engine: NDVI and Satellite Image Collections — vegetation-index calculation.
- Google Earth Engine: Sentinel-2 Cloud and Shadow Masking — a major source of false changes.
- Google Earth Engine: Change Detection and Clustering — a technical worked study.
- How AI Recognises Objects in Photos — understanding image categories.
- How Convolutional Neural Networks Read Images — how spatial features are learned.
- How AI Is Tested Before Release — independent checks.
- How AI Works: Complete Overview — the wider series.