An Image-Based Machine Learning Method for Urban Features Prediction with Three-Dimensional Building Information

Machine learning has been proven to be a very efficient tool in urban analysis, using models trained with big data. We have seen research that applies a generative adversarial network (GAN) to train models, feeding the street map and visualized urban characteristics to predict certain urban features. However, in most cases, the input map is a two-dimensional (2D) map that only stores the land type data (e.g., building, street, green space), hence reducing building information to only the ground-floor area. The identities of buildings with similar floor areas can be hugely different, which may contribute to the prediction errors in previous machine-learning models. In this research, we emphasize the importance of the use of an image-based neural network to analyze the relationship between urban features and the constructed environment. We compare the model that uses traditional street color maps as the input set, against a new input set with more detailed building data. Once trained, the model with the enhanced input set yields output at a higher level of accuracy in certain areas. We apply the new model framework to three selected urban features predictions: rental price, building energy cost, and food sanitary ratio. A broad range of new research could be conduct with our new framework. © 2023 and published by the Association for Computer-Aided Architectural Design Research in Asia (CAADRIA), Hong Kong.