Decomposing urban heat islands at the micro-scale: A hybrid physics-inspired and data-driven modelling framework

High-resolution monitoring of canopy-layer Urban Heat Island (UHI) dynamics remains challenging because fixed stations are sparse and satellite products do not directly measure near-surface air temperature. This study presents a hybrid neural decomposition model for estimating sub-daily urban–rural air-temperature differences ( Δ T ) from opportunistic vehicle-mounted measurements. The model decomposes Δ T into a cell and time-conditioned UHI potential, U H I p ( c , t ) , a non-negative meteorological modulation term, U H I m ( t ) , derived from a 92-hour reference-station history and an empirical correction for mobile-sensor airflow bias. Applied to Nicosia, Cyprus, it used approximately 700,000 one-second measurements collected by municipal vehicles during August–September 2025 across 863 retained 200 × 200 m cells spanning sampled road corridors. Comparison with six third-party weather stations in sampled cells yielded station-level Mean Absolute Error (MAE) values of 0.25–1.09 ° C . It estimated a transition from morning cool-island conditions (−0.7 to −1.3 °C) to afternoon urban warming (＋2.4 to + 3.1 ° C ), followed by positive nocturnal differences (＋1.4 to + 1.6 ° C ). Learned U H I p was higher in compact low-rise Local Climate Zones (LCZs) and lower in open low-rise zones. Associations with Normalised Difference Vegetation Index, building fraction, elevation and LCZ class supported a plausible, though incomplete, interpretation of spatial structure. Low model-estimated U H I m values during the 9–11 August heatwave occurred under limited route coverage and should not be interpreted as evidence of a general heatwave mechanism. Overall, opportunistic municipal sensing can support micro-scale UHI characterisation along sampled road corridors, although neighbourhood-scale, seasonal and cross-city validation remains necessary.