Large-scale Digital Surface Models (DSMs) can be generated cheaply from satellite imagery through stereo-photogrammetry, but the resulting 3D maps frequently suffer from noise, outliers and voids. Aerial LiDAR, by contrast, delivers highly accurate elevation data at considerably greater cost.
A new research paper proposes bridging that gap with generative AI. The authors study diffusion models conditioned jointly on photogrammetric DSMs and Pléiades satellite imagery to refine vertically co-registered elevation maps.
The team introduces a modified Stable Diffusion 3 architecture featuring a pruned text stream and a patch-wise normalization strategy. According to the paper, this design enables stable training on LiDAR data and supports transfer from natural images to elevation maps.
Experiments conducted across French cities show that multimodal conditioning improves elevation accuracy. In the in-context cities, Dense Urban RMSE fell from 6.00 to 3.45 metres, while in the held-out city of Bordeaux it dropped from 4.16 to 2.77 metres.
The work sits alongside a growing cluster of related efforts on diffusion-based elevation super-resolution, LiDAR diffusion and geo-aware foundation models for Earth observation.