GAN-Generated Elevation Models in Computational Fluid Dynamics: A Feasibility Study for Complex Urban Terrain

Recently developed methods to simulate very high-resolution (VHR) wind fields over complex urban terrain rely on high-quality three-dimensional vector representations of building information. Unfortunately data of that kind is sparsely available on a worldwide scale. In this work, we investigate the applicability of computational fluid dynamics (CFD) on 2.5D digital surface models (DSMs) automatically generated by generative adversarial network (GAN) from globally available satellite data which includes photogrammetric DSMs and pan-chromatic (PAN) images. The obtained results demonstrate that the GAN-based DSMs are reasonable alternatives to rarely available level of detail 2 (LoD2) vector data, promoting large coverage, continuous wind field derivation over complex terrain.