Lunar image registration

Every pixel, in its true place.

Sub-pixel registration for lunar imagery that holds across Sun angle, scale and viewpoint.

Sources
PRADAN · LRO · SELENE · USGS
Models
DINOv2 · SuperPoint · LightGlue
Runs
Fully offline

001 / 006

A controlled map stays true.

Decades of orbital imagery have been stitched into mosaics tied to laser-altimeter ground control. At the south pole that map is accurate to metres — it is the fixed reference everything else is measured against.

002 / 006

A new image arrives in the wrong place.

A fresh high-resolution image is placed on the Moon using predicted orbit and pointing. Small errors in those predictions move it away from the ground it actually shows — close enough to look right, too far to be trusted.

003 / 006

Sun angle changes everything.

Near the pole the Sun never climbs far above the horizon. The same crater throws a different shadow on every pass, so two images of the same ground can share almost no pixels that look alike.

004 / 006

Matching finds the same ground.

Contrast is normalised, then learned features propose tie-points: places where the new image and the map show the same ground. When shading defeats keypoints, foundation-model embeddings and area correlation take over.

005 / 006

Verification rejects what disagrees.

Not every tie-point is right. RANSAC keeps only those that agree on one geometric transform, then checks they spread across the whole frame instead of clustering in one corner.

006 / 006

Registered.

The verified transform moves the image onto the map's grid. It is exported as a GeoTIFF that opens on the right ground in any GIS, alongside a report of how accurate the fit is.