DUNE: A Machine Learning Deep UNet++-based Intrinsic Forecast Approach to Monthly, Seasonal and Annual Climate Forecasting

Abstract

Capitalizing on the recent availability of ERA5 monthly averaged, long-term data records of mean atmospheric and climate fields derived from high-resolution reanalysis, deep- learning architectures provide an alternative to physics-based daily numerical weather predictions for subseasonal-to-seasonal (S2S) and annual forecasts. A novel Deep UNet++-based Ensemble (DUNE) neural architecture is introduced, incorporating encoder-decoder structures with residual blocks. When initialized with data from the prior month, season, or year, this architecture delivers an AI-based global forecasts for monthly, seasonal, and annual means of T2m and SST. ERA5 monthly mean data are utilized as inputs for T2m over land, SST over oceans, and climatological monthly solar radiation at the top of the atmosphere, covering 40 years of data to train the model. Validation forecasts are conducted for another two years, followed by five years of forecast evaluations to capture natural annual variability. Rigorous testing was performed using a cross-validation approach with multiple validation and testing periods. The DUNE-trained inference weights enable forecasts to be generated within seconds. Performance metrics such as RMSE, ACC, and HSS are analyzed globally and across specific regions. DUNE AI’s forecasts outperform persistence, climatology, and multiple linear regression across all domains. DUNE forecasts demonstrate comparable statistical accuracy to NOAA’s operational monthly outlooks for the U.S., but at significantly higher spatial resolutions. RMSE seasonal comparisons with NOAA’s NMME and ECMWF’s SEAS5 show that DUNE outperforms both in most seasons and captures major anomalies with finer spatial detail.