New Quantitative Approach for the Morphological Similarity Analysis of Urban Fabrics Based on a Convolutional Autoencoder

Urban fabric similarity analysis and classification represent a useful analysis framework in many urban studies. Classical approaches include qualitative descriptions or manually selected indicators, depending largely on the researcher's knowledge. This paper proposes a new method for extracting concise and integrated quantitative indicators to represent urban fabrics via deep learning. Figure-ground images are taken as training data for a convolutional autoencoder (CAE) model, and the outputs of the neurons in the bottleneck layer of the CAE are extracted as the compressed feature vectors (CFVs) to represent the plots. Then, the plots can be compared, clustered and visualized based on these CFVs. In this study, 345 residential plots of Nanjing, China are taken as samples to demonstrate the modeling process, clustering and visualizations. The results show that the CFV is an effective indicator to represent urban fabrics. The generation of the CFV takes into consideration both statistical and geometrical features, with the latter normally described qualitatively as patterns. CFV consists of several aspects of urban fabric without the need to balance the weights. The CFV can serve as a basis for further urban development interpretations, morpho-typology studies and other social, economic and microclimate studies that have relationships with the urban fabric.