GANmapper’s innovations allow it to generate realistic and high-resolution images compared to other AI models that appear very artificial. The model is based on the concept of Generative Adversarial Networks (GANs). The motivation is to create building footprints on an image to show the exact urban morphology of the city and generate maps from it. GANmapper aims to solve the problem by generating reality-true photos with angular and scale fidelity. This simultaneously opens several doors in fields such as the film industry, video game development, medicine and many more.
The project is based on the 2022 article by Abraham Wu and Filip Biljecki “GANmapper: geographical data translation”. They present a new method for creating spatial data using a generative adversarial network. By using available geodata to create maps with more information.
In our project we use the source code of GANmapper and other input data to produce maps for the city of Hamburg. For this we use available street networks and building data as input. This can be particularly useful in locations where detailed and high-resolution data is currently lacking. The quality of the results depends on the city shape and the selected zoom factor. The basic idea is to compare three differently densely populated regions in northern Germany. The question to be answered is whether the model predicts the building footprints in the same way. Furthermore, the quality of the building footprints will be qualitatively assessed by means of additional measurements. The evaluation of the accuracy is based on a comparison of the building footprints with existing maps.
The result achieved is not definitive, because the goals were only partially reached. However, the potential of the model becomes clear. To achieve better results, there is a need to optimise the parameters. By changing the values of the parameters, new settings can be tried and the results perfected.