Comparative Evaluation of Multi-Source Geospatial Data and Machine Learning Models for Hourly Near-Surface Air Temperature Mapping

Accurate estimation of hourly near-surface air temperature (NSAT) is critical for climate analysis, environmental monitoring, and urban thermal studies. However, existing temperature datasets remain constrained by coarse spatial resolution and limited hourly accuracy. This study systematically evaluates four widely used land surface temperature (LST) datasets—MODIS, ERA5-Land, FY-2F, and CGLS—and five machine learning models (RF, MDN, DNN, XGBoost, and GTNNWR) for NSAT estimation across two contrasting regions in Shaanxi, China: a complex-terrain region in southwestern Shaanxi and the urban area of Xi’an. Results demonstrate that single-source LST inputs outperform multi-source LST stacking, largely due to compounded systematic biases across heterogeneous datasets. MODIS provides the best performance in the mountainous region, while CGLS excels in the urban environment. Among all models, GTNNWR—which explicitly captures spatiotemporal non-stationarity—consistently achieves the highest accuracy, reducing RMSE by 44.8% and 44.2% relative to the second-best model in the two study areas, respectively, whereas the remaining four models exhibit broadly comparable performance. This work identifies effective data–model configurations for generating high-resolution hourly NSAT products and provides methodological insights for climate and environmental applications in regions with complex terrain or strong urban heterogeneity.