High-Resolution Daily Air Temperature Estimation over China: An Explainable Stratified Stacking Ensemble Approach

Near-surface air temperature (NSAT) is crucial for climate and hydrological studies, making the development of accurate estimation models along with high-resolution datasets essential. Although remote sensing (RS) technology combined with machine learning (ML) has been widely applied to the retrieval of land surface variables, several challenges remain in estimating NSAT: (1) the optimal algorithm for different land cover types still unclear; (2) the differences in the contributions of various predictors to NSAT across different land cover types are not well understood; (3) there is still a lack of high-resolution, spatiotemporally continuous NSAT datasets. To address these, we developed an NSAT estimation framework for mainland China using 831 meteorological stations and nine RS predictors, integrating stacking ensemble learning with SHAP analysis. We systematically compared stratified modeling by land cover type (StraM) with integrated modeling (IntM). Results show: (1) NSAT estimation accuracy varies by ML algorithm and land cover type. StraM with stacked optimal algorithms achieved the best performance (NSATmin: RMSE = 2.647 °C; NSATave: RMSE = 1.883 °C; NSATmax: RMSE = 2.681 °C); (2) The StraM method consistently achieved significantly higher estimation accuracy across all land cover types compared to the IntM. (3) The contributions of predictors to NSAT estimation varied among land cover types, with land surface temperature remaining the most important variable. The generated NSAT dataset aligns with existing datasets and ground observations in capturing the spatiotemporal characteristics of NSAT over China, while demonstrating improved capability in representing spatial details. This study provides a robust foundation for developing representative NSAT products applicable to ecological monitoring and climate research. • A new 1-km daily air temperature dataset (NSATDC) was developed across China. • Land-cover stratified stacking ensemble (StraM) improves estimation accuracy. • NSATDC provides seamless daily mean, max, and min temperatures for all surfaces. • SHAP analysis validates the physical consistency of the learned mechanisms.