**1. Introduction**

Urban morphology refers to the multidisciplinary study of urban forms regarding the physical environment, the cultural preservation process, and sustainable development [1,2]. The morphological approach provides the idea that morphology has the potential to be an animating force for urban design [3]. Many studies relate urban typo-morphology to studies of the citizens' lives, the social economy, and the energy system efficiency [4–6]. Therefore, urban morphology provides a valuable basis for urban planners and managers. Urban morphology is related to complex urban system analysis, such as energy performance [7], citizen behavior, and economic benefits [8].

On the other hand, in the urban renewal process, communications with the as-built urban fabric need to be considered in urban design. In a study of urban neighborhoods [9],

**Citation:** Cai, C.; Guo, Z.; Zhang, B.; Wang, X.; Li, B.; Tang, P. Urban Morphological Feature Extraction and Multi-Dimensional Similarity Analysis Based on Deep Learning Approaches. *Sustainability* **2021**, *13*, 6859. https://doi.org/10.3390/ su13126859

Academic Editor: Tan Yigitcanlar

Received: 23 May 2021 Accepted: 11 June 2021 Published: 17 June 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

evolutionary patterns related to sustainable urban neighborhoods were extracted based on the morphological classification and clustering of footprint patterns over time.

In addition to simulation-based urban morphological analysis [10], researchers focusd on data-driven approaches for urban morphological studies recently by considering the methods for constructing relations between the as-built and the to-be-built environment, such as multi-dimensional (e.g., the geometric dimension and social dimension) of urban forms. Technologies, such as 3D scanning, depth detection, multi-directional scanning, and simultaneous localization and mapping (SLAM) are becoming increasingly sophisticated. The abundance of web-based map data, such as AutoNavi, Baidu, and OpenStreet Map (OSM), allows architects to grasp data efficiently. In the information abundance, machine learning approaches support design by providing designers with previous cases for new design solutions based on the case-based reasoning (CBR) [11].

In data-driven urban design, learning from reality (in this case, urban morphology) helps decision makers in making comprehensive design decisions and researchers in the study of spatial form-related functionality and performance. On one hand, the suggestions for urban morphology design in the decision-making process take social, environmental, and economic factors into consideration. The suggestions serve as a reference as well as guidance for decision-makers.

Urban morphology is the physical carrier of quality of life. In the urban design process, the lack of a concrete understanding of urban morphology by non-professional decisionmakers limits the space for discussion by designers [12]. Both developers and designers could develop discussions and ideas from the cases in similar situations, introducing the information and knowledge. On the other hand, the associations of urban morphology with functionality and further consequences supports researchers in morphology-related studies. The suggestions for urban cases could be based not only on the spatial form but also on the morphology-related traffic networks, energy performance, economic conditions, and so on, thus, supporting further scientific utility.

Effective morphological quantification methods for cases is crucial in terms of the data representation for a case retrieval system. However, cities are developed in complicated historical, economic, and behavioral contexts. Every city is unique in urban form [13]. There is not a clear-cut answer of the critical factors or the factor weights for city development or collapse simulation [14]; therefore, deduction and verification methods are challenging to apply in a comprehensive urban morphology study for urban design. For example, in the MApUCE tools chain study, a processing chain was proposed to calculate 64 standardized urban morphological indicators to represent the buildings, blocks, and spatial units [15].

More indicators could be extracted for more precise representation. However, each indicator's weight influences the calculation since it should be appropriate for reaching the global fitting of the instances, and there were still missing factors by selecting and calculating indicators. Moreover, the critical indicators varied from the cases from different cultural and historical contexts, which need to be studied and verified. Quantifying the morphology of a large number of cases with indicators would lead to a generalizability limitation. Therefore, a descriptive framework for urban support analysis is essential, and its effectiveness could be shown in the subsequent data analysis and visualization of urban case retrieval.

In developing an efficient approach for urban morphological quantification methods for similarity analysis, three facets of obstacles are observed: (1) the construction of multi-dimensional urban dataset that includes geometrical and social information, (2) the quantification of urban morphological features with concise but informative descriptions, and (3) similarity calculations of the extracted morphological features. According to the above discussion, an automatic data mining method would support the detailed information collection of the related infrastructures from various aspects. A fully automated feature extraction method may help to overcome the drawbacks of manually selecting indicators and balancing weights. A feature extraction method considering the statistical and overall characteristics of the instances could be introduced.

Machine learning approaches help to efficiently represent and retrieve cases from a considerable amount of data [16]. Deep learning is one of the branches of machine learning. Deep learning algorithms promote evolutionary methodologies for morphological analysis. This approach is robust with pictorial datasets because of the development in convolutional neural networks [17]. The convolutional methods support a fully automated feature extraction process among a large amount of data. This represents the samples' characteristics with concise and comprehensive information in feature vectors. Methods, such as image-data-based (RGB), numerical labeling, and semantic segmentation, are dedicated to feeding samples into neural networks with continuous and informative features. Cluster analysis supports comparison and similarity studies by data-deduction and distance calculating techniques, taking the feature extraction data as a basis.

We clarified our study scope on case retrieval to support design decision-making and serve as a basis for further scientific utility via deep learning, as, regarding the information abundance and complexity of cities, we need a solution space implying urban knowledge rather than certain answers. For example, the efficient similarity analysis of urban associations (e.g., the morphology, traffic, energy, and economy) represents the task-oriented retrieved cases for decision makers in certain applications and researchers for further scientific analysis, based on texts, images, models, and other representation media carrying concrete urban information.

Therefore, the construction of the spectrum of cases based on situation similarity is a promising way to efficiently introduce references from the information abundance for a wider discussion space for design decision making and precise association of morphology with urban consequences for scientific studies. The reference's effectiveness depends on the case quantification methods and the related social information (e.g., infrastructure, industrial distribution, and traffic conditions) of the cases. The method would also have potential for a general search engine in terms of urban morphology. Therefore, an effective morphological quantification approach and multi-dimension datasets, including related geometric and social information, would be needed for comprehensive urban design decision-making.

In this study, we propose a multi-dimensional similarity analysis approach for introducing cases in similar situations from the information abundance. The similarity analysis includes morphological similarity and social situation similarity. The proposed method combines data mining and cluster analysis via deep learning, taking the residential cases in Nanjing, China, for instance. In this study, a multi-dimension dataset, including geometrical information and infrastructure information, is constructed. The samples' morphological features are extracted into high-dimensional feature vectors (HDFV) via a deep convolutional neural network. This study further completes the case retrieval based on the HDFV. The architects can retrieve cases according to the plot-shape-similarity or building-distribution-similarity, along with the infrastructure information. The significance of the study is as follows:

