High-resolution pedestrian-level air temperature evaluation in tropical high-density cities–Using machine learning and point‑based morphology

• Public-data-driven ML model evaluates pedestrian-level air temperature • Point-based morphology surpasses site-averaged indicators in temperature evaluation • Diurnal variability dominates over monsoon effects on model accuracy • Final model achieves high accuracy (R² = 0.99, MSE = 0.23°C², error < 1°C) • Scalable, real-time framework deployed in an interactive microclimate platform Rapid urbanization and climate change have intensified thermal stress in high-density tropical cities, where pedestrian-level air temperature is shaped by atmospheric conditions and three-dimensional morphology. This study develops a high-resolution, rapid, and generalizable model to estimate pedestrian-level air temperature using only publicly accessible meteorological data and GIS-derived morphological indicators. To capture fine-scale spatial variability, a new point-based indicator called the surrounding building height-to-distance ratio (H/D) is introduced and integrated with conventional site-averaged metrics. Using Singapore as a representative tropical city, six machine-learning algorithms were benchmarked to identify the best-performing approach. Two modelling strategies were then compared: a year-round model and microclimate-specific models stratified by monsoon season and diurnal cycle. Results indicate that nighttime models (trained on 7 PM-7 AM data) achieve the highest accuracy, followed closely by the year-round model (trained on full annual data), while daytime models (7 AM-7 PM) demonstrated slightly weaker generalization. Consequently, the final framework employs nighttime models during 7 PM-7 AM and the year-round model during 7 AM-7 PM, with diurnal variability exerting a much stronger influence than seasonal monsoon differences. The final model achieved R² = 0.99 and MSE = 0.23°C², with typical evaluation errors below 1°C. SHAP analysis highlights strong morphology-meteorology interactions and shows that point-based indicators outperform site-averaged metrics in shaping local air temperature. Practically, the 1-meter, minute-scale model rapidly generates continuous thermal maps capturing fine-grained urban heterogeneity. Integrated into the Microclimate Digital Platform, it enables on-demand pedestrian-level temperature prediction, providing a scalable and data-efficient tool for real-time thermal assessment and heat-resilient planning in tropical high-density cities.