The fast, cheap way to build 3D maps from satellites has been making beautiful models that are geometrically wrong. If you want a 3D model of a city, the shape of the land and every building on it, the best raw material is often satellite photos. They cover huge areas cheaply. The catch is that a satellite is a terrible photographer for this job. A drone can circle a building and study it from every side, but a satellite screams past hundreds of kilometres overhead and grabs each shot from almost the same steep angle. Worse, the handful of useful pictures of any one place were taken on different days, sometimes months apart. Different seasons, different sun, shadows sprawling in different directions across the same rooftop. Feed those mismatched views into a 3D reconstruction and the errors compound. The ground that comes out is gently warped, like a map printed on a sheet that's been left in the sun. For pretty pictures that's fine. For flood modelling, where a few centimetres of fake slope sends the simulated water the wrong way, it's a real problem. There are two modern ways to build 3D from photos. One is meticulous but slow and gobbles memory. The other is a newer trick that paints a scene as millions of tiny coloured fuzzy blobs, and it's gloriously fast and light. When people first pointed that blob method at satellite images, the results looked great and measured badly. Lovely renders, wonky geometry. A team at Ohio State, in a method they call SatSplat, went after the geometry instead of the looks. They swapped the round blobs for flat little discs that hug surfaces more honestly, taught the system the strange near-parallel way a satellite sees the ground, and let it nudge each camera's assumed position to fix small pointing errors. Then they tackled the light directly, mapping where every shadow falls and recolouring each date's image so the same roof stops reading as two different objects. On two standard benchmarks for this exact task, SatSplat cut the average height error of the reconstructed surface by 11.93% against the previous blob-based satellite methods. That's the difference between a model you can run an engineering calculation on and one you have to apologise for. It did this while using 31% less peak graphics memory, and that saving is what lets you build these models over whole cities instead of single blocks. This is an incremental, but pretty good step in the direction of every geospatial nerd's dreams: 3d models built from satellite images. link to full article: arxiv.org/abs/2606.28581