Machine-learning reconstruction of ERA5-Land temperature at local administrative units

Extreme climate impacts are often managed at neighborhood and local administrative scales, yet reanalysis products remain spatially coarse and are not aligned with administrative boundaries such as Census Block Groups (CBGs). This study develops a machine-learning-based change-of-support reconstruction workflow to translate hourly ERA5-Land 2-m air temperature (2000–2024) from its native approximately 9-km grid to CBG areal support across Maryland. ERA5-Land fields were mapped to CBGs using sparse bilinear weights and decomposed into a leave-one-year-out hourly climatology and CBG-specific anomalies. Month-specific Random Forest models reconstructed ERA5-Land-derived anomalies using temporal lags, harmonic terms, dew point, 10-m winds, and static spatial predictors. Validation against 14 independent daily station records showed annual root-mean-square error (RMSE) values of 1.2–3.1 °C and Pearson correlations of 0.95–0.99, with larger errors in winter and high-elevation CBGs. A supplementary 2024 subdaily consistency check using four independent NOAA/NCEI airport stations and 34,471 common station-hours yielded a pooled hourly RMSE of 1.86 °C and correlation of 0.983, with larger errors during nighttime than daytime hours. Before bias adjustment, relative to deterministic ERA5-to-CBG mapping using nearest-neighbor and inverse-distance weighting, the reconstructed series had lower RMSE at 11/14 and 12/14 stations, respectively; after leave-one-year-out bias adjustment, the corresponding counts were 10/14 and 9/14. In unadjusted seasonal comparisons, RMSE reductions reached 14% relative to nearest-neighbor mapping and 6.84% relative to inverse-distance weighting, although overall gains were modest. Targeted sensitivity analyses showed that temporal-memory predictors dominated model importance, whereas contemporary landscape variables and direct geographical coordinates contributed comparatively little. The resulting hourly CBG-support temperature series provides administrative-unit climate information for ecological, heat-risk, public-health, and climate-adaptation research; applications that depend strongly on local hourly extremes should include additional application-specific validation.