Improving Subseasonal to Seasonal Global 2-m Temperature and Precipitation Forecasts Using a Deep Learning–Based Postprocessing Approach
- 1 March 2026
- journal article
- Published by American Meteorological Society in Monthly Weather Review
- Vol. 154 (3) , 391-405
Abstract
Numerical weather prediction models continue to exhibit flaws in forecasting subseasonal to seasonal (S2S) 2-m temperature (T2M) and total precipitation (TP) beyond 2 weeks. In this study, a deep learning–based postprocessing model for the S2S probabilistic forecast of global T2M and TP was developed using a residual U-Net (ResU-Net) algorithm. The model was developed based on deterministic input information from weekly reforecasts of the European Centre for Medium-Range Weather Forecasts (ECMWF); these aggregates included five upper-air atmospheric variables at three pressure levels (250, 500, and 850 hPa) and six surface variables. ResU-Net produced global probabilistic tercile forecasts of T2M and TP for weeks 1–6, including biweekly aggregates of weeks 3–4 and 5–6. Compared with the ECMWF S2S forecasts, the ResU-Net probabilistic forecast demonstrated superior probabilistic forecasts for up to 6 weeks. The most important predictors to forecast T2M in lead time weeks 1 and 2 are above mean sea level (MSL), 850-hPa geopotential height (Z850), 850-hPa temperature (T850), and T2M, while weeks 4–6 are sea surface temperature (SST), 200-hPa geopotential height (Z200), and 200-hPa temperature (T200). For TP forecast, TP and total column water are important predictors during weeks 1–3 and TP and SST are important predictors for weeks 4–6. The ocean–atmosphere coupling and large-scale teleconnections influence longer lead week forecasts of T2M and TP. Therefore, combining numerical weather prediction models and deep learning show promising S2S probabilistic forecasts.Keywords
Funding Information
- Korea Meteorological Administration (RS-2024-00403698)
- National Research Foundation of Korea (No. RS-2023-00301938)
- National Research Foundation of Korea (No. RS-2023-00301938)