Deep learning for multi-year ENSO forecasts.

Area of Science:

• Climate Science
• Artificial Intelligence
• Oceanography

Background:

• El Niño/Southern Oscillation (ENSO) variations cause significant global climate extremes and ecosystem impacts.
• Accurate long-lead ENSO forecasts are crucial for effective policy and resource management.
• Current forecasting models struggle with lead times exceeding one year.

Purpose of the Study:

• To develop a highly skillful ENSO forecasting model with extended lead times.
• To leverage deep learning for improved prediction accuracy and mechanism analysis.
• To overcome limitations of existing dynamical forecast systems.

Main Methods:

• Utilized a convolutional neural network (CNN) with transfer learning for model training.
• Trained the CNN on historical climate simulations and reanalysis data (1871-1973).
• Validated the model's performance using the Nino3.4 index and sea surface temperature data (1984-2017).

Main Results:

• The CNN model achieved skillful ENSO forecasts up to 1.5 years in advance.
• Demonstrated significantly higher correlation skill for the Nino3.4 index compared to state-of-the-art dynamical models.
• Accurately predicted the zonal distribution of sea surface temperatures, a key ENSO characteristic.
• Heat map analysis confirmed the use of physically plausible precursors for ENSO prediction.

Conclusions:

• Deep learning, specifically CNNs, offers a powerful new approach for ENSO forecasting.
• The developed CNN model surpasses existing systems in prediction accuracy and lead time.
• This model provides a valuable tool for understanding ENSO mechanisms and improving climate predictions.