A technical report describing DEMGAN, a Conditional Image-to-Image GAN approach that translates global low-resolution (30m) DEMs plus OpenStreetMap vector layers into synthesized high-resolution (2m) DEMs. The model uses ResNet-based generators and PatchGAN discriminators, trained on paired 30m–2m data (50,000 pairs) from sources like CopernicusDEM, USGS, Vancouver City Government, and SwissTopo. Results show substantially lower FID and RMSE versus bicubic and CycleGAN baselines, and experiments indicate some cross-city generalization (Pittsburgh, Vancouver, Denver, Zurich).
Questions this document answers
50,000 pairs of data
CopernicusDEM (30m)