Human activities have profoundly altered urban surface properties, resulting in pronounced spatiotemporal heterogeneity in urban heat exposure. Near-surface air temperature ( T a ) characterizes key aspects of the thermal environment experienced by humans and is therefore an important variable in urban climate assessment. However, accurately mapping the diurnal spatial distribution of T a at the urban scale remains highly uncertain under sparse observational conditions, particularly in hot-humid regions characterized by frequent cloud cover and strong land-atmosphere interactions. To address this challenge, we propose a data-efficient multiscale data fusion framework for urban Ta spatial prediction. The framework employs reconstructed land surface temperature as a core input, integrates urban features across multiple neighborhood scales, and uses TabPFN to model the relationship between optimal predictors and station-based Ta observations. Evaluation across eight diurnal periods in a hot-humid city during summer demonstrates that the proposed approach enables accurate and robust Ta mapping at a spatial resolution of 70 m under both clear-sky and cloudy conditions. 10-fold cross-validation yields RMSEs of 0.45–0.90 °C, MAEs of 0.35–0.72 °C. Compared with existing methods that rely on long-term, dense observations or single-scenario assumptions, the proposed framework achieves comparable or superior performance for all-sky diurnal urban Ta spatial prediction.
Luo et al. (Wed,) studied this question.