Scenario-Specific Contrast Enhancement for Satellite Co-registration

Summary

When you want to detect changes between two satellite images of the same location, those images first need to be perfectly aligned — a process called co-registration. That's harder than it sounds: the position of the sun, the season, the angle at which a satellite photographs the scene, and the so-called parallax effect (which makes tall buildings, for example, appear shifted between two photos of the same spot) all make it difficult to find reliable correspondences between images.

My research tests whether adding a contrast-enhancing preprocessing step — CLAHE (Contrast Limited Adaptive Histogram Equalization) — can improve the performance of MatchAnything, a state-of-the-art AI model for image matching. Since no dataset exists with satellite images where the exact geometric transformation between captures is known, I built a proxy experiment: starting from the public AID dataset (10,000 satellite images across 30 categories such as desert, beach, forest, and residential areas), I generated my own controlled transformations and noise, so the quality of every match could be measured objectively.

The results showed that for 9 of the 30 categories, a specific CLAHE setting produced a significant improvement compared to no CLAHE — for 7 of those, in both match quality and success rate. The Pond category benefited most clearly: a high clip limit prevented the otherwise featureless water surface from being incorrectly matched with vegetation. For categories with little inherent structure, such as desert and bare land, a low clip limit worked best instead. For most of the remaining (often highly repetitive or uniform) categories, CLAHE brought no benefit and sometimes even a disadvantage.

The conclusion: the hypothesis that an optimal CLAHE setting can be determined per category is partially confirmed. Contrast enhancement is not a universal solution but a scenario-specific one: which setting helps depends strongly on the visual characteristics of the landscape.

Download

Note: the full thesis document itself is written in Dutch.