Deep Learning-Driven Morphological Dataset and Analysis Methods for Chinese Campuses
Citation
Wang Y, Wang X, Tang P. Deep Learning-Driven Morphological Dataset and Analysis Methods for Chinese Campuses[C]// GAME CHANGER? Planning for just and sustainable urban regions - AESOP Annual Congress proceedings- Vol. 36 (2024), Paris, France, 7-12 July 2024.
Abstract
Modern campuses in China display distinct morphological characteristics, evolving to form unique patterns as subsystems within the urban environment. Hence, the approaches for comprehensive analysis for those urbanized Chinese campus morphology (UCCM) are important. This study proposes a framework for dataset construction and morphology recognition of UCCM, using visual representing learning methods. Computer vision technologies are used to acquire the morphology patches of 1257 campuses. We analyse the campus morphology with our proposed multi-dimensional morphometrics. Then, we constructed multiple morphological cluster maps for UCCM in terms of road, building and landscape, respectively. The cluster maps show significant compliance with human visual perception. Compared with classic morphometrics, our approach excels in learning implicit morphological characteristics with lower data processing demands and less reliance on expert experience.
Keywords
Urbanized Chinese campus morphology, Morphometric, Visual Representation, Self-organizing Map, Unsupervised learning
1.Data Acquisition
There is a significant missing for geometric data of campuses in China, especially those located in suburbs or new towns, on widely-used open map platforms like open Street Map(OSM). To address this issue, we propose a method to extract information from map tiles acquired from online map platforms. The exact map tiles for campus are pinpointed and crawled with their geometrical information. The imagery data are then converted to GeoTiff format and clipped by campus boundary.
2.Morphometrics for Campus
Referring the method of urban morphometric, we summarize the metrics corresponding to each dimension as well as their calculations and the morphology element data required for the analysis.38 Campuses in Nanjing are selected to practice the analysis approach applying morphometrics (Figure 6). The binary image data of the various elements in the dataset are vectorized and assigned with attributes for the analysis.
3.Visual Representation Learning
Nonlinear dimension reduction methods, which have been proven to be efficient for high dimensional features, including UMAP, TSNE and Autoencoder, are tested separately to visualize extracted features from imaginery data. The results exhibit that the campus form gradients rather than falling into simplistic categories, which can be explained by the gradient distribution of both the size and form of Chinese campus. Therefore, compared with classic linear clustering methods like k-means, SOM is applied to achieve continuous visualization as a spectrum in a 2D space.
