ABSTRACT
This study presents a novel two-step algorithm for mapping hourly urban air temperature at a 250 m spatial resolution in Seoul, addressing the limitations of satellite imagery and ground-based observation stations for monitoring urban heat island (UHI) effects. The algorithm involves downscaling 1 km MODIS land surface temperature (LST) data to 250 m using a machine learning Random Forest model and estimating hourly air temperature from the downscaled data. To mitigate cloud contamination issues in satellite LST, a method for combining LST data from various time periods is proposed. Furthermore, we introduce a multi-task learning estimation model for simultaneously calculating 24-hour hourly air temperatures. The proposed technique demonstrates high estimation accuracy with an average RMSE of approximately 1.0 ℃ across all 24 hours. The air temperature maps effectively simulate detailed urban temperature patterns and their variations over time. Analyzing thermal environment distribution using Local Climate Zones (LCZs), we identify distinct UHI patterns in Seoul, with higher building density and taller structures within open built-up type LCZ classes correlating to higher temperatures at night. The hourly air temperature maps offer valuable insights into urban thermal environments and the multi-task learning approach holds promise for practical applications.
