Answers are grounded in this paper and cite their sources.
Near-surface air temperature (Ta) is a vital indicator for urban heat island (UHI) research, but continuous high-resolution Ta data remain scarce. Land surface temperature (LST) is frequently used as a Ta alternative, though spatiotemporal distinctions are rarely evaluated. We investigated spatiotemporal heterogeneity, nonlinear correlations, and influential factors of time-aligned Ta and LST in Hong Kong during summers and winters (2018–2023). Integrating meteorological station data and satellite imagery, a machine learning (ML) framework and SHapley Additive exPlanations (SHAP) analysis evaluated the contributions of nineteen meteorological, urban morphological, topographic, landscape, and anthropogenic factors. Both temperature metrics show heterogeneity under day/night and seasonal contexts. Aligned Ta-LST pairs were positively correlated, but fitting outcomes varied across local climate zones, limiting LST-to-Ta linear translation accuracy, especially in summer. ML models showed acceptable accuracy for Ta (RMSE: 0.53–1.14 °C) and LST (RMSE: 1.08–2.70 °C) estimation. Temporal–spatial SHAP decomposition found that coarse-resolution reanalysis variables had nearly all predictive power in temporal signals rather than in spatial gradients. Target-demeaned attribution isolated spatial drivers: building height was a leading contributor to nighttime Ta; impervious surface fraction influenced daytime LST. SHAP analysis reflected several findings: daytime shading suppresses surface heating, while nighttime radiative trapping within deep canyons sustains elevated Ta. Leave-one-station-out validation confirmed spatial generalizability of the models. These results suggest that UHI mitigation is most effective when strategies target each metric separately. Reducing LST alone may do little for ambient heat stress if the 3D morphological and aerodynamic processes that control urban air temperature are left unaddressed.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: