Developing a CNN-based, block-scale oriented Local Climate Zone mapping approach: A case study in Guangzhou

Abstract

The limited available urban land in megacities is represented by large areas of impervious underlying surfaces and buildings, which are expressed by the basic urban fabric unit of the “city block.” The concept of the Local Climate Zone (LCZ) scheme is capable of quantifying the combined influences that different urban elements have on block-scale local climates, rendering block-scale LCZ classification and block-scale-oriented LCZ map production crucial for climate-conscious urban planning and design. Taking the city of Guangzhou as a case study, deep learning was combined with the Convolutional Neural Network (CNN) to perform effective image recognition, vector cutting, and automatic area statistics based on multilevel road networks. Road network division, block area processing, characteristic parameter calculation, and LCZ classification were integrated to automatically generate block-scale-oriented LCZ maps, which was achieved via threshold contrast using the recommended parameter intervals of the calculated characteristic parameters. Basic land surface temperature (LST) data for summer to winter in Guangzhou were then applied, and correlation analysis was conducted between the block-scale LST vector and LCZ characteristic parameters. The results demonstrate the applicability and potential of the CNN algorithm for urban climate analysis. The developed CNN-based LCZ mapping approach prioritizes “block” scale analysis and typical characteristic parameters in each block are obtained with high calculation accuracy. This study contributes to guiding parameterized climate-sensitive urban planning.