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

TL;DR

The paper tackles subseasonal-to-seasonal and annual climate forecasting by replacing physics-only daily NWP with a data-driven Deep UNet++ Ensemble (DUNE) trained on ERA5 monthly means. DUNE leverages a multi-encoder–decoder UNet++ with residual and dense skip connections to predict $T_{2m}$ over land and $SST$ over oceans, using anomaly-based training relative to a 1950–1979 climatology and a moving-window scheme for extended horizons. Across monthly, seasonal, and yearly means, DUNE achieves lower $RMSE$, higher $ACC$, and competitive $HSS$ relative to persistence, climatology, and even NOAA’s forecasts, at a high 0.25° global resolution with rapid inference enabling ensemble forecasts. The work demonstrates practical potential for AI-assisted climate prediction, with implications for wildfire risk assessment and near-term climate services, and provides data and code access for reproducibility. DUNE’s approach—anomaly training, integrated ocean-land fields, and intrinsic ensembles—offers a scalable path toward high-resolution, fast S2SA climate forecasting.

Abstract

Capitalizing on the recent availability of ERA5 monthly averaged long-term data records of mean atmospheric and climate fields based on high-resolution reanalysis, deep-learning architectures offer an alternative to physics-based daily numerical weather predictions for subseasonal to seasonal (S2S) and annual means. A novel Deep UNet++-based Ensemble (DUNE) neural architecture is introduced, employing multi-encoder-decoder structures with residual blocks. When initialized from a prior month or year, this architecture produced the first AI-based global monthly, seasonal, or annual mean forecast of 2-meter temperatures (T2m) and sea surface temperatures (SST). ERA5 monthly mean data is used as input for T2m over land, SST over oceans, and solar radiation at the top of the atmosphere for each month of 40 years to train the model. Validation forecasts are performed for an additional two years, followed by five years of forecast evaluations to account for natural annual variability. AI-trained inference forecast weights generate forecasts in seconds, enabling ensemble seasonal forecasts. Root Mean Squared Error (RMSE), Anomaly Correlation Coefficient (ACC), and Heidke Skill Score (HSS) statistics are presented globally and over specific regions. These forecasts outperform persistence, climatology, and multiple linear regression for all domains. DUNE forecasts demonstrate comparable statistical accuracy to NOAA's operational monthly and seasonal probabilistic outlook forecasts over the US but at significantly higher resolutions. RMSE and ACC error statistics for other recent AI-based daily forecasts also show superior performance for DUNE-based forecasts. The DUNE model's application to an ensemble data assimilation cycle shows comparable forecast accuracy with a single high-resolution model, potentially eliminating the need for retraining on extrapolated datasets.

Figures (26)

• Figure 1: The DUNE AI Architecture
• Figure 2: (a) Residual Block (b) Regular CNN Block
• Figure 3: Global RMSE of Monthly Mean Forecasts: (a) Persistence (PYSM), (b) Climatology, and (c) DUNE AI Method. The comparison shows that the DUNE AI model consistently achieves lower RMSE values, indicating superior forecasting skills. Notably, during the record warm year of 2023, the DUNE AI model exhibited relatively small variability compared to the baseline models. Over the five years, the DUNE AI model displays slightly higher variance in January, February, and December but maintains a uniformly small variance for the remaining months.
• Figure 4: Global ACC of Monthly Mean Forecasts: (a) Persistence (PYSM), (b) Climatology, and (c) DUNE AI Method. The comparison shows that the DUNE AI model consistently achieves higher ACC values, indicating superior forecasting skills. Notably, during the record warm year of 2023, the DUNE AI model exhibited relatively small variability compared to the baseline models. Over the five years, the DUNE AI model displays higher ACC value of 0.8 compared to 0.5 (baselines) in addition to maintaining a smaller variance.
• Figure 5: Visualization of forecast errors for December 2023: (a) Persistence (PM), (b) Persistence (PYSM), (c) Climatology, (d) DUNE AI Method. RMSE and ACC values are displayed below each plot. (a) and (b) highlight significant prediction errors using the Persistence method over North America and Eurasia. In contrast, the DUNE AI model demonstrates substantial performance improvements, particularly in the El Niño region, South America, Africa, North America, and Boreal regions across North America and Eurasia.