Weakly-Supervised Change Detection in Satellite Imagery via Adversarial Class Prompting

Imagine comparing two satellite photos of the same area taken at different times and asking a very picky, well-trained inspector to highlight only the meaningful changes (like new buildings or destroyed infrastructure), even though nobody ever labeled those changes by hand. This method teaches the AI to become that inspector using only coarse, cheap labels and a clever ‘good cop / bad cop’ game inside the model so it learns what real change looks like versus noise.