A Rapid Preliminary UrbanLayout Method 

Based on GauGAN and Automatic Dataset Construction

| Keywords |

Machine learning, GauGAN, Urban Layout ,Urban morphological features,Automactic data construction

| Abstract |

Recently Machine learning (ML) has been widely used in the urban layout process. Many studies have demonstrated that training the Generative Adversarial Network (GAN )model by labeling enables machines to learn urban morphological features and urban layout logic. However, the number and quality of datasets greatly determine the quality of supervised ML results. Manually selecting urban samples that meet the criteria to create a dataset and labeling it by urban morphological features is a tedious and repetitive process. There is still a lack of a hybrid method to automatically produce urban samples datasets instead of manual work to train a customized ML model for an efficient urban layout process. We demonstrate a method that integrates a Gau GAN framework with an automated sample collection system and automatic label system. The system is implemented in HOUDINI modeling software. The sample collection system automatically collects samples from the OSM open-source map platform in a set-up region. The samples are filtered to meet designer criteria such as density, height, functional diversity, etc. Label system automatically extracts representative urban morphological features such as boundary, void, openness, etc. Then the system labels urban morphological features on the bitmap to make the dataset. The aim is to train an ML model for fast multiple urban layouts and visualize suggestions by designers simply drawing labels about urban morphological features. Casestudy demonstrated the system automatically collects and labels a 2000 sample dataset in Milan. The trained model shows an efficient urban layout capability for different plots in Milan. And the model can assist the designer to generate a variety of schemes by different label inputs in the same site. Then some limitations are discussed in the automatic label sessions. This framework has the potential to become an urban morphology syntactic generative tool in the future.

| Figures |