*2.3. Deep Learning for Morphological Analysis*

Compared with conventional methods, the advantages of the deep learning method is that it is an end-to-end (e.g., image to segmentation label) process, with the potential of connecting task-oriented things [32]. Today, convolutional neural networks (CNNs) are one of the most prominent deep learning approaches for image processing and computer vision. In a convolution neural network (CNN), the feature mapping of the images extracts the input's underlying features by convolution kernels, promoting deep learning algorithms in recent decades [33]. Features of the input are continuously convolving into multidimensional vectors layer by layer. Neural networks have been implemented for solving problems in the field of architecture, such as the prediction of energy performance [34], pattern recognition of 2D images [35], as well as typological form-finding on 3D models [36].

New data analysis techniques have been proposed for the design process to address the abundance of variables. Neural network techniques, such as classification, prediction, and cluster analysis, provide technical support for regression models for data analysis. Cluster analysis is a powerful technique for urban morphological analysis, which extends the technical support of high efficiency and accuracy for classification and clustering analysis from a large amount of data.

In a study of typo-morphology in Lisbon [37], the block and street types were classified based on the prepared plans using the k-means clustering algorithm. The ETH team's "City of indexes" project interpreted urban morphological patterns based on the massive data of city images combined with personal preferences [38].

Concise but informative feature vectors extracted via deep learning were used to quantitatively characterize morphological features [39]. With the introduction of deep learning, architects need to focus on formulating inputs related to architectural problems, understanding the meaning of each layer's output, and finding the application sceneries. Since cluster analysis is an unsupervised technique, it excludes labeling data, which is timeconsuming. It automatically features the images with high-dimensional data to support further analysis.
