What is satellite change detection?

The idea in one sentence

Satellite change detection compares two or more observations of the same location, acquired at different times, to identify where the surface appears different.

The output is usually a map: a mask or heatmap showing where change is likely, often turned into polygons that a person can review, measure, and share.

Common approaches

Methods range from simple arithmetic on pixels to deep learning models trained on pairs of images.

• Image differencing and ratios: subtract or divide the same band across two dates and threshold the result.
• Index comparison: compare spectral indices such as NDVI for vegetation or water indices for surface water.
• Post-classification comparison: classify each date into land-cover classes, then compare the class maps.
• Learned models: neural networks that take two dates as input and predict where meaningful change occurred.
• Time-series methods: model the normal behavior of each pixel over many dates and flag departures from it.

Why different pixels are not always real change

Many things make two images of the same place look different without anything having happened on the ground. A good workflow screens for them before raising an alert.

• Seasonal and crop-cycle change, such as green-up, harvest, or leaf fall
• Clouds, haze, and cloud shadows in optical imagery
• Different sun angles, which change shadows and brightness
• Small misalignment between images, which creates false edges
• Comparing different sensors with different bands and resolutions
• Water level changes from tides or normal seasonal flows

Observed change versus inferred cause

Imagery can show that a surface changed from vegetation to bare ground. It usually cannot prove why. Construction, fire, harvest, and clearing can look similar from orbit.

That is why GeoVeyra keeps the observed change separate from any interpretation, and routes consequential findings to a person for review.

What good change detection looks like

Whatever the method, a trustworthy result carries its context with it.

• The exact observations compared, with acquisition dates and sensor
• Notes on cloud cover, coverage gaps, and resolution limits
• A boundary for each finding, not just a pin
• A review status, so everyone knows what has been confirmed