UHI-Bench: Benchmarking Dual-Source Urban Heat Island Modeling Across Cities in Diverse Climate Regimes

Urban heat islands (UHIs) are a critical climate risk, yet researchers often conflate two distinct heat signals: land surface temperature (LST-UHI) and near-surface air temperature (AirT-UHI). This paper introduces UHI-Bench, the first benchmark designed to model both sources simultaneously. By integrating dynamic meteorological drivers (e.g., wind, cloud cover) with static urban morphology (e.g., building density, road networks) across 20 cities and nine climate classes, the authors provide a standardized framework for evaluating UHI imputation, forecasting, and cross-city generalization.

The authors organized UHI-Bench around a signal-mechanism-transfer framework. They first analyzed the consistency between LST-UHI and AirT-UHI, finding that they capture different physical aspects of heat and are not interchangeable. They then evaluated over 20 baseline models—ranging from classical machine learning to time-series foundation models—across five tasks. A key focus was testing how well models trained in one city can predict UHI in another, specifically investigating whether climate-diverse training sets outperform larger, climate-homogeneous sets.

The study reveals that no single model family is uniformly superior, though foundation models demonstrate consistent stability across diverse tasks. Environmental covariates (meteorology and urban form) significantly improve performance, but their utility is highly task-dependent: meteorological data are crucial for LST-UHI imputation, while static urban features are more effective for AirT-UHI. Most importantly, the authors found that cross-city transferability is better explained by the overlap in UHI regimes—such as diurnal patterns and heat amplitude—than by traditional Köppen climate labels. This suggests that future urban heat modeling should prioritize selecting source cities based on their thermodynamic similarity to the target city.