Multi‐Source Data‐Driven Downscaling and Bias Correction of ERA5 ‐Land Near‐Surface Air Temperature Over Complex Terrain

ABSTRACT

Near‐surface air temperature exhibits local spatial differences in complex terrain, and the native spatial resolution of ERA5‐Land near‐surface air temperature cannot directly meet application requirements. Taking Sichuan Province as the study area, this study developed a statistical downscaling and bias‐correction workflow for daily ERA5‐Land near‐surface air temperature based on observations from 6045 stations during 2020–2024, ERA5‐Land 2 m air temperature, high‐resolution land‐surface predictors, and ERA5/ERA5‐Land Reanalysis‐Derived Auxiliary Predictors (RDAPs). The performance of Elastic Net (EN), Bayesian‐Optimized Random Forest (BO‐RF), and Random‐Search‐based Long Short‐Term Memory network (RS‐LSTM) under different combinations of model and predictor set was compared across multiple evaluation dimensions. The results showed that all three models performed better after RDAPs were introduced. BO‐RF with RDAPs and RS‐LSTM with RDAPs were the two best‐performing combinations, with the lowest RMSE values for T _avg , T _max , and T _min on the independent final validation set of 0.94°C, 1.40°C, and 1.12°C, respectively. The two best‐performing combinations also showed relatively stable monthly RMSE reductions, with no evident month‐to‐month fluctuations in accuracy; for all three temperature variables, the proportions of validation stations with RMSE improvement exceeded 97%. Grouped Permutation Feature Importance (GPFI) further indicated that the near‐surface humidity and surface thermal state group and the pressure–wind–circulation background group were the RDAP predictor groups with relatively high contributions, whereas the terrain group was the main contributing group among the high‐resolution land‐surface predictors. However, the error reduction of the best‐performing combinations remained limited for sporadic extreme samples.