Predicting land surface temperature by different climate classification methods: A case study of Singapore

Abstract. The urban thermal environment has become a challenge to humans in consideration of rapid urbanization and global warming. Various climate classification methods have been developed to analyze urban form and the urban heat island phenomenon. However, there is a lack of cross-comparison studies carried out to examine the accuracy of predicting land surface temperature by different climate classification methods (local climate zone, urban functional zone, and hybrid zone that integrates the strengths of local climate zone and urban functional zone), as well as their performance in statistical and machine learning models (ordinary least squares regression, geographically weighted regression, and random forest regression). Accordingly, this study focuses on comparing the performance and accuracy of predicting land surface temperature via different climate classification methods. In addition, the relative importance and marginal effect of factors on land surface temperature are discussed based on the approach with the highest accuracy. The results show that: random forest model performs best in predicting land surface temperature (average R²: 0.72); hybrid zone is the most accurate approach to predict land surface temperature (R²: 0.84); and urban functional zone (R²: 0.80) performs slightly better than local climate zone (R²: 0.76). This study helps urban planners and designers to assess which climate classification methods can more accurately predict and explain the influence of urban form on land surface temperature, and provides some insights into urban design strategies to improve the thermal environment.