*3.3. Cluster Analysis and Visualization*

Using image data of the residence plots as the input for clustering is often impractical due to the gigantic size of the data matrix converted from the images. Therefore, a process of dimension reduction is often required. Currently, there are three mainstream techniques for data reduction: Principal Component Analysis, the t-SNE algorithm, and an Auto-Encoder. In this experiment, we used the t-SNE algorithm to map high-dimensional feature points to a two-dimensional plane without losing the information of the feature vectors. The samples with similar features were placed as neighbors in the cluster cloud (Figure 6). Figure 6 shows the spectrum of all the cases based on morphological similarity. The left picture represents the clustering results in terms of the cases' plot shapes, while the right picture shows the cases' plots with buildings. The more similar the cases are, the closer they are on the clustering map.

**Figure 6.** Atlas of the clustering analysis of 3817 residential plots in Nanjing, China.

To show the clustering map more clearly, we zoomed in to some parts and present them in Figure 7. The samples are shown on the same scale. Clusters of squaring, narrow, or irregular shapes can be intuitively seen in Figure 7a. The result is different in Figure 7b, where the distribution of buildings influences the clustering result. Different residential types could be observed, such as plots with few rows of buildings, closely spaced residential buildings, and loosely arranged villas, etc. We can intuitively see that the plots belonging to the same cluster have morphological features in common. The cluster analysis performed better in near-square plot shapes rather than irregular plot shapes, as more cases had square plots.
