Advancing S2S and Decadal Forecasts with DUNE: A Deep UNet++ Ensemble Approach

Abstract

Introduction •The DUNE (Deep UNet++ Ensemble) model leverages deep learning to enhance climate forecasting. •Capable of predicting land and sea temperatures over monthly, seasonal, and annual timescales. •Offers rapid and reliable forecasts, outperforming traditional Numerical Weather Prediction (NWP) models. Key Features and Performance •Continuously updates predictions over a 12-month horizon, improving long-term reliability. •Demonstrates enhanced performance when trained on longer-term averages. •Future tests aim to extend forecasting capabilities to decadal scales. Comparative Advantages •Outperforms baseline methods (persistence, climatology, linear regression) across multiple regions. •Matches NOAA’s forecast accuracy while providing higher spatial resolution (0.25° vs. 2°). •Achieves significantly faster computation, making it suitable for operational forecasting. Impact •Highlights the potential of deep learning to advance climate forecasting. •Provides accurate, high-resolution predictions addressing key climate and weather science challenges.