Incorporating planning intelligence into deep learning: A planning support tool for street network design

With the emergence of deep learning techniques, procedural and example-based modeling have been increasingly applied to support automatic content generation and visualization for planning decisions (Hartmann et al., 2017). Procedural modeling relies on manually designated rule sets to produce proposals. Parish and Müller (2001) made one of the first attempts to generate three-dimensional city models for visualization using procedural approaches, where a Lindenmayer system was used to grow road networks and buildings conditioned on global goals and local constraints. Given an initial and a final road point, Galin et al. (2010) developed a cost minimization function to automate path creation, considering the slope of the terrain and natural obstacles. The function was then extended to generate hierarchical road networks between towns at a regional level (Galin et al., 2011). Similar procedural principles can also be applied to allocate land use, subdivide blocks and generate buildings (see, e.g., Chen et al., 2008; Lyu et al., 2015). In comparison, example-based approaches learn from real-world cases in a preprocessing step to extract features and adopt them as templates. Hartmann et al. (2017) developed an automatic road generation tool, StreetGAN, using a generative adversarial network (GAN) to synthesize street networks in a fix-sized region that can maintain the consistency of urban layouts learned from the training data set. Similarly, Kempinska and Murcio (2019) trained Variational Autoencoders (VAEs) using images of street networks derived from OpenStreetMap to capture urban configurations using lowdimensional vectors and generating new street networks by controlling the encoded vectors.