Cities are now central to addressing global changes, ranging from climate change to economic resilience. There is a growing concern of how to measure and quantify urban phenomena, and one of the biggest challenges in quantifying different aspects of cities and creating meaningful indicators lie in our ability to extract relevant features that characterize the topological and spatial patterns of urban form. Many different models that can reproduce large-scale statistical properties observed in systems of streets have been proposed, from spatial random graphs to economical models of network growth. However, existing models fail to capture the diversity observed in street networks around the world. The increased availability of street network datasets and advancements in deep learning models present a new opportunity to create more accurate and flexible models of urban street networks, as well as capture important characteristics that could be used in downstream tasks. We propose a simple approach called Convolutional-PCA (ConvPCA) for both creating low-dimensional representations of street networks that can be used for street network classification and other downstream tasks, as well as a generating new street networks that preserve visual and statistical similarity to observed street networks.
