The rapid development of artificial intelligence (AI) has led to numerous advances in recent years and has become a part of our daily lives. The areas of application of AI are manifold. The main one is to automate complex tasks. For one, it can be used in the film industry to create realistic sets and special effects, or in video game development to create graphics and animations. AI is able to interact and process information. But it is not only used in the entertainment industry; doctors can also generate realistic 3D models of organs to help them diagnose and treat diseases. Today, AI is even used in the financial sector for automated analysis of financial data and market trends. It enables advanced autonomous systems in robotics used in industry, transport and space. AI is also used in image and speech recognition, language translation, security and agriculture. This shows that AI has virtually unlimited applications and the potential to improve many aspects of our daily lives.
One particular model is the GANmapper. GAN stands for Generative Adversarial Network and the motivation behind the development lies in the geographical and spatial application. The basis for the project is the article by Abraham Wu and Filip Biljecki “GANmapper: geographical data translation”, which describes how spatial data can be automatically created in maps using the model. This paper introduces GANmapper, which should be able to create realistic-looking and hierarchically correct building floor plans from other datasets, such as road networks.
Our project takes up the GANmapper model and tests it. It is investigated whether the GANmapper can generate building floor plans in maps for the greater Hamburg area from available geodata. A region with a high population density is chosen. Likewise, a region in Hamburg with a low population density is selected for comparison. The aim is to generate complete and correct maps with scale and angular accuracy. The model is attempted to be tested on available geodata, such as building and road data in its topology. Buildings are chosen as input data because they are conspicuous and can be easily checked for completeness and accuracy. At the end, the individual results of the different regions are compared with each other. It is specifically investigated here whether the model predicts the building footprints in the different regions in the same way.
In order to answer the key question the existing model is first adapted for the new data sources. adapted. The model is trained with the new training data on population density, road network and building footprints and the results are examined at the end. Based on the shape, size and location of the building footprints created, a judgement can be made about the result.
The article is structured as follows. Section 2 “Generative Adversarial Networks” provides an introduction to GAN. It gives a brief overview, explains how it works and what it can be used for. Section 3 “GANmapper” includes only the GANmapper and gives the basics. Section 4 “Data” introduces the data used and explains how it is processed. Section 5 “Methods and Results” presents how our idea was implemented and deals with the results. Section 6 “Discussion” discusses the results, methods and data used. The project is concluded with section 7.