*4.3. Discussions on the Method*

The advantages of the proposed method regarding similarity analysis can be discussed by comparing the study using conventional methods, including the Roma urban renewal project mentioned in Section 2.2.

First, the HDFV had higher efficiency on the urban fabric quantification and similarity analysis. In the study of Roman, the authors extracted several geometric indicators to evaluate the block shape similarity. This took time for the extraction and verification process of selecting the indicators and balancing each weight. The HDFV carries comprehensive morphology characteristics by matrix operation of the sample pixels based on the deep learning model. It saves effort in addressing a large number of images since the whole process from feature extraction to case retrieval is performed automatically.

Secondly, the HDFV has more generalizability in applying the method to various urban morphologies. The Roman study had a limitation when evaluating the building distribution texture similarity. It required another extraction and verification process of selecting indicators to describe the building distribution characteristics, which is a more complicated and diverse process compared with describing the block shape. With the HDFV, we could perform the similarity investigation of plot shape or building distribution under the same framework. The proposed method goes beyond the morphology types from different historical contexts because it learns from the samples directly and clusters them based on the feature vector distance.

Thirdly, the proposed method has more flexibility regarding similarity-based case retrieval. In the study of Roman, the system recommended another case that was similar to the input. However, we could find a series of cases that are similar to the input from sufficient samples. The result space would be broader if there were more samples in the dataset. In this way, more references and information could be brought to designers.

What is more, the proposed method could broaden the similarity analysis because it takes related social information (e.g., the infrastructure) into consideration by integrating a multi-dimensional dataset. In addition to the morphology and the POI information studied in this experiment, the framework could be implemented for quantification and similarity analysis with more information related to the urban sustainability. For instance, the cultural background, the energy performance, the traffic conditions, etc. could be added to the dataset, depending on the design task. The image-based similarity analysis can be done via the deep learning model. This would provide precise references in similar situations to better support the designers' decision-making towards sustainable cities.

The limitations of this study involve that the highly automated process increases the difficulty of emphasizing the specialties from a particular aspect. There are three main limitations. First, the insufficient number of samples with similar plot sizes leads to some noise in the case retrieval results. The results would be more robust if there were around 1000 cases with similar plot sizes. This limitation could be overcome by simply adding more cases collected from cities around Nanjing based on AutoNavi.

Second, all 3D information (e.g., building height and building shape) is lost since the samples are represented as 2D images for learning. This drawback could be overcome by adding one more color channel to represent the building height or by using voxels instead of pixels to describe the plot in three dimensions. Third, the target needs to be trained together with the cases in the dataset. In other words, once a new target is introduced, the entire neural network has to be retrained.
