GANmapper: geographical data translation

We present a new method to create spatial data using generative adversarial network (GAN). Our contribution uses coarse and widely available geospatial maps of less features at the finer scale in built environment, bypassing their traditional acquisition techniques (e.g. satellite imagery or land surveying). In work, we employ use road networks as input generate building footprints conduct experiments 9 cities around world. The method, which implement tool release openly, enables translation one dataset another with high fidelity morphological accuracy. It may be especially useful locations missing detailed high-resolution those that are mapped uncertain heterogeneous quality, such much OpenStreetMap. quality results is influenced by urban form scale. most cases, suggest promising performance tends truthfully indicate locations, amount, shape buildings. work has potential support several applications, energy, climate, morphology studies areas previously lacking required inpainting regions incomplete data.