ABSTRACT
In this chapter, we explore the application of Convolutional Neural Networks (CNNs) for the mapping of Local Climate Zones (LCZs), an innovative approach to objectively quantify the urban heat island effect across diverse urban and non-urban landscapes. Drawing a comparison between traditional LCZ mapping methodologies – including manual, GIS-based, Earth Observation (EO)-based, and hybrid techniques – and CNNs, we highlight the advantages of the latter. Our unique data processing technique involves using a moving window of size 32*32 at intervals of 32 pixels. This method demonstrates a marked improvement over traditional methodologies in terms of performance. In an effort to achieve a balanced representation of diverse LCZ types, we employ data augmentation strategies on each polygon, which yield effective results. We provide a detailed evaluation of several renowned CNN models – VGG, ResNet, DenseNet, and GoogLeNet – for LCZ mapping in Beijing. Each model is trained using pre-set parameters, a learning rate of 0.001, and a 100-epoch cycle. Our findings reveal overall accuracy scores ranging from 80% to 90% among the models, with GoogLeNet and DenseNet121 exhibiting particularly impressive performance. In conclusion, this chapter underscores the potential of utilizing CNNs in LCZ mapping and sets the stage for future advancements in urban climate studies.
