**3. Materials and Methods**

Figure 1 shows the general workflow of our study. First, we collected web map data, including geometrical information and Point of Interests (POI) in Nanjing, which were downloaded with the open source AutoNavi API. We filtered the collected data into a geometrical dataset and additional social information (distribution of related infrastructure) based on functionality with the ArcGIS software. We exported images for the case slices in terms of plots.

Second, the morphological features were automatically extracted through a deep convolutional neural network with inception-v3 modules into high dimensional feature vectors (HDFV). Third, the cluster analysis was visualized based on the t-SNE algorithm in a two-dimensional plane. The Euclidean distance was applied to calculate the similarity between the cases. Finally, the performance of the HDFV on the similarity analysis was verified by a comparison case retrieval study.

**Figure 1.** The general workflow of quantitative urban morphological analysis for case-based design via deep learning.
