Yu et al. (2025) Spatio‐Temporal Network With Self‐Attention Mechanism for Improved ENSO Prediction

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Earth and Space Science
• Year: 2025
• Date: 2025-11-27
• Authors: Nan Yu, Changhong Hu, Jinghuan Wang, Weibo Rao, Siwei Liu, Minghui Yang, Gang Chen, Jinze Li
• DOI: 10.1029/2024ea004179

Research Groups

[Information not available from the abstract.]

Short Summary

This study proposes ACTNet, a novel deep learning model combining Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and a self-attention mechanism, to improve long-lead prediction of the Niño 3.4 index and classify different ENSO types.

Objective

• To develop a new neural network architecture (ACTNet) for improved long-lead prediction of the Niño 3.4 index by effectively capturing spatiotemporal dependencies in global climate variables.
• To classify different ENSO types (e.g., Eastern Pacific, Central Pacific El Niño and their La Niña counterparts) using deep learning based on historical observed SST anomalies.

Study Configuration

• Spatial Scale: Global (processing global sea surface temperature, heat content, zonal wind, and meridional wind).
• Temporal Scale: Input data covers the past 12 months; predictions are generated at a monthly resolution up to 24 months in advance.

Methodology and Data

• Models used:
  • ACTNet (Convolutional Neural Network + Long Short-Term Memory + Self-Attention mechanism) for Niño 3.4 index prediction.
  • Conventional CNN and CNN + LSTM models (for comparison).
  • LSTM model for ENSO type classification.
  • SINTEX-F, CanCM4 (mentioned as typical physically coupled models for context).
• Data sources:
  • Global sea surface temperature (SST).
  • Global heat content.
  • Global zonal wind (UA).
  • Global meridional wind (VA).
  • Historical observed SST anomalies (specifically Niño 3 and Niño 4 indices for ENSO type classification).

Main Results

• ACTNet successfully predicts the Niño 3.4 index up to 24 months in advance, achieving correlation coefficients exceeding 0.5.
• ACTNet demonstrates improved spatiotemporal feature extraction and superior long-lead prediction skill compared to conventional CNN and CNN + LSTM models.
• An LSTM model achieved a classification accuracy of 70.5% for six defined ENSO types at a 12-month lead time.

Contributions

• Introduction of ACTNet, a novel deep learning architecture that integrates CNN, LSTM, and a self-attention mechanism, for enhanced ENSO prediction.
• Demonstration of improved long-lead ENSO prediction skill (up to 24 months) for the Niño 3.4 index using ACTNet, surpassing conventional deep learning models.
• Successful application of deep learning for multi-type ENSO classification, providing a valuable tool for understanding regional climate impacts.

Funding

[Information not available from the abstract.]

Citation

@article{Yu2025SpatioTemporal,
  author = {Yu, Nan and Hu, Changhong and Wang, Jinghuan and Rao, Weibo and Liu, Siwei and Yang, Minghui and Chen, Gang and Li, Jinze},
  title = {Spatio‐Temporal Network With Self‐Attention Mechanism for Improved ENSO Prediction},
  journal = {Earth and Space Science},
  year = {2025},
  doi = {10.1029/2024ea004179},
  url = {https://doi.org/10.1029/2024ea004179}
}