Change detection that survives the seasons
Difference two satellite passes of the same field and the result lights up everywhere. Snow fell, a shadow swung, a crop turned from green to gold. None of it is the change anyone tasked, and all of it buries the change they did.
The noise is periodic, and that is the lever
Most false positives come from predictable cycles: sun angle, vegetation, water level, weather. They are large and everywhere, but they repeat. A model that knows the expected seasonal state for a place and date can flag departures from that, instead of from one stale reference frame.
What actually gets compared
- Co-register first. Align to sub-pixel accuracy so a pixel is compared with the same ground sample, not its neighbour.
- Structure over brightness. A new edge or a filled void outranks a radiometric shift; geometry is harder to fake than colour.
- Tune to recall. Suppress too hard and you miss the event, so every release is swept against a labelled set before it ships.
The target is not zero alerts. It is that every alert is worth an analyst's time, with the before and after frames attached so the call takes seconds.