Development of LCZ mapping method combining GIS with machine learning

Local Climate Zones (LCZ) have been used worldwide in the field of urban climate. In most previous studies, LCZ maps—zoning maps for LCZ categories—were generated using the WUDAPT Level 0 method with satellite images and machine learning techniques. As this method can generate maps without relying on geometric information about urban morphology, it is useful for urban climate analysis in cities where such numerical information is not publicly available. However, this method involves the subjectivity of the mapmaker, resulting in different threshold values between LCZ categories for different mapmakers, which makes it difficult to compare and analyze urban climates within the same category across multiple cities generated by different researchers. Although one solution is the use of Geographic Information System (GIS) data to determine the LCZ category for each location, some LCZ categories cannot be determined based on numerical information alone. This study proposes a method for generating LCZ maps that allows all LCZs to be classified according to their actual conditions and matched to the correct urban morphological values by combining GIS and machine learning techniques. Using the proposed method in this study, LCZ maps were generated for Tokyo, Hong Kong, and Singapore. As a result, the consistency of the average building height and building coverage ratio with the reference values was verified. For all form-based categories (LCZ 1-6, 9) in the three cities, average building height (ABH) and building coverage ratio (BCR) were found to be within the range of the reference values. Additionally, the Weather Research and Forecasting (WRF) simulation incorporating the map across multiple Asian cities showed that the closer a city is to the equator, the greater the effect of sea breezes in reducing future temperatures, and identified the influence of different latitudes on future changes in open high-rise areas.