A Novel Classification Framework for Urban Heat Island Analysis Using Unsupervised Clustering

Rapid urbanization and climate change are intensifying urban heat islands (UHIs), necessitating precise analysis and effective urban planning. While previous studies have examined UHI intensity based on built morphology using Local Climate Zones (LCZs) at individual city scales or across Köppen-Geiger climate zones, research on the spatial configuration of LCZs within and between cities remains relatively limited. This study proposes a novel classification framework to identify climate zones at intra- and inter-city levels, which are not adequately captured by existing Köppen-Geiger climate classifications or the element level of LCZs themselves. To achieve this, we computed UHI intensity across global cities using ERA5-Land air temperature data. We extracted latent space representations from the UHI time series using an autoencoder and applied K-means++ clustering to categorize these features into five distinct classes. The proposed classification effectively captured variations in UHI intensity both within and between cities. Notably, the moderate UHI class, characterized by an annual mean UHI close to zero, exhibited the lowest median proportion of built-type LCZs (65%) and the highest proportion of land cover types (34%). The clustering results revealed spatial patterns distinct from the Köppen-Geiger climate classification, which overlooks variations in UHI effects across cities within the same climate zone. By providing a refined categorization of UHI intensity, this study enhances our understanding of intra-city climate variations and offers a valuable framework for UHI mitigation strategies, climate-sensitive urban planning, and sustainable city development.