Evaluating ERA5 Correction Methods in the Brazilian Cerrado Using Spatial Validation and Predictor Similarity

Five methods for correcting daily ERA5 air temperature and relative humidity in the Brazilian Cerrado were evaluated against INMET observations during 2001–2024 using spatial and spatiotemporal validation. The primary assessment included 99 temperature stations and 98 humidity stations. For temperature, monthly linear scaling and linear regression reduced the median station-level root mean square error (RMSE) by 8.08% and 7.77%, respectively. Linear scaling was retained as an interpretable regional monthly adjustment, without a claim of statistical superiority over regression. Performance varied geographically: its median station RMSE changes ranged from a 2.55% deterioration to a 14.36% reduction across five blocks. Spatial-weight sensitivity showed similar positive residual dependence for both methods. Relative humidity gains were small and lacked consistent concordance improvements, so no regional correction was applied. Temperature correction produced 23,510,412 cell-day estimates over 2682 cells, decreasing mean temperature from 25.17 to 24.69 °C. A temperature-and-seasonality similarity screen included 96.19% of these estimates. At held-out stations, median dissimilarity had a near-zero association with RMSE (Spearman ρ = 0.026), limiting its interpretation as a reliability measure. The original constant-width 3.877 °C error interval describes pooled residual dispersion. A separate nested spatial calibration achieved 90.56% overall coverage, but coverage fell to 84.13% in one block and 85.48% in the warmest training-defined temperature stratum. Regional mean improvements therefore coexist with geographic and climatic limits to transferability. Results are based on daily means screened using an 80% hourly-completeness requirement (at least 20 distinct valid hours per station–variable–UTC day).