Comparative Assessment of Post-Processing Approaches for GEFSv12 Summer Maximum Temperature and Heat Extreme Forecasts Over CONUS

The increasing frequency of summer time heat extremes across the continental United States highlights the need for reliable forecasts of daily maximum temperature, especially at longer lead times. This study evaluates three post-processing approaches: the Detrended Quantile Mapping (QM), Artificial Neural Networks (ANN), and a sequential Hybrid ANN–QM technique, to improve sub-seasonal summer maximum temperature (T_max) and heat-extreme forecasts from the NOAA NCEP Global Ensemble Forecast System version 12 (GEFSv12) over the continental United States (CONUS) during 2000–2019. Forecasts were verified against CPC observations using deterministic, categorical, and probabilistic metrics across multiple lead times. Raw GEFSv12 forecasts reproduced large-scale T_max patterns but exhibited significant regional biases and declining skill at extended lead times. While QM effectively corrected systematic biases, it provided limited improvements in forecast–observation association. Conversely, ANN significantly enhanced forecast skill by reducing RMSE and better preserving interannual variability. The Hybrid approach synthesized the strengths of both, leveraging ANN for nonlinear error correction and QM for distributional calibration, consistently yielding the most balanced performance and highest skill for heat-extreme prediction. These findings underscore the efficacy of combining machine learning and statistical post-processing to advance sub-seasonal temperature forecasting.