Classification and mapping of urban canyon geometry using Google Street View images and deep multitask learning

Abstract

Urban canyon classification plays an important role in analyzing the impact of urban canyon geometry on urban morphology and microclimates. Existing classification methods using aspect ratios require a large number of field surveys, which are often expensive and laborious. Moreover, it is difficult for these methods to handle the complex geometry of street canyons, which is often required by specific applications. To overcome these difficulties, we develop a street canyon classification approach using publicly available Google Street View (GSV) images. Our method is inspired by the latest advances in deep multitask learning based on densely connected convolutional networks (DenseNets) and tailored for multiple street canyon classification, i.e., H/W-based (Level 1), symmetry-based (Level 2), and complex-geometry-based (Level 3) classifications. We conducted a series of experiments to verify the proposed method. First, taking the Hong Kong area as an example, the method achieved an accuracy of 89.3%, 86.6%, and 86.1%, respectively for the three levels. Even using the field survey data as the ground truth, it gained approximately 80% for different levels. Then, we tested our pretrained model in five other cities and compared the results with traditional methods. The transferability and effectiveness of the scheme were demonstrated. Finally, to enrich the representation of more complicated street geometry, the approach can separately generate thematic maps of street canyons at multiple levels to better facilitate microclimatic studies in high-density built environments. The developed techniques for the classification and mapping of street canyons provide a cost-effective tool for studying the impact of complex and evolving urban canyon geometry on microclimate changes.

Files

Additional details