Peng et al. (2025) Assessing seasonal prediction of DGPI over the Western North Pacific by the Climate Forecast System and its improvement using deep learning

Identification

• Journal: Atmospheric Research
• Year: 2025
• Date: 2025-11-19
• Authors: Shanchi Peng, Chao Wang, Liguang Wu, Haikun Zhao, Jian Cao
• DOI: 10.1016/j.atmosres.2025.108628

Research Groups

• State Key Laboratory of Climate System Prediction and Risk Management/Key Laboratory of Meteorological Disaster of Ministry of Education/Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science and Technology, Nanjing, China
• Pacific Typhoon Research Center and Earth System Modeling Center, Nanjing University of Information Science and Technology, Nanjing, China
• Department of Atmospheric and Oceanic Sciences and Institute of Atmospheric Sciences, Fudan University, Shanghai, China

Short Summary

This study evaluates the Climate Forecast System's (CFS) seasonal prediction skill for the Dynamic Genesis Potential Index (DGPI) in the Western North Pacific, finding limited skill in operational forecasts, and then significantly improves this prediction using Convolutional Neural Network (CNN) models trained with CMIP6 and reanalysis data.

Objective

• To evaluate the performance of the Climate Forecast System (CFS/CFSv2) in seasonally predicting the Dynamic Genesis Potential Index (DGPI) over the Western North Pacific (WNP), focusing on interannual variability in both spatial distribution and basin-wide mean values.
• To develop and assess Convolutional Neural Network (CNN) models to improve the seasonal prediction skill of DGPI, addressing the limitations of current operational models.

Study Configuration

• Spatial Scale: Western North Pacific (WNP)
• Temporal Scale: Seasonal prediction; evaluation period 1990–2022, with sub-periods 1990–2010 (CFS Reforecast) and 2011–2022 (operational CFSv2 and CNN).

Methodology and Data

• Models used: Climate Forecast System (CFS/CFSv2), Convolutional Neural Network (CNN)
• Data sources: CFS Reforecast predictions (1990–2010), operational CFSv2 forecasts (2011−2022), Coupled Model Intercomparison Project Phase 6 (CMIP6) historical simulations (for initial CNN training), reanalysis data (for CNN transfer learning and refinement), observed DGPI (for evaluation).

Main Results

• CFS exhibits limited skill in capturing DGPI spatial variability, with a mean spatial correlation of 0.21 during 1990–2022.
• CFS Reforecast predictions (1990–2010) capture interannual variability well in basin-mean DGPI (correlation coefficient, r = 0.71), but operational CFSv2 forecasts (2011−2022) perform poorly (r = 0.19) due to an inability to reproduce mean state changes.
• CNN models significantly improve DGPI prediction skill, achieving a mean spatial correlation of 0.85 for anomaly fields during 2011–2022.
• CNN models show a strong correlation with observed basin-mean DGPI (r = 0.88), markedly outperforming CFSv2.

Contributions

• Provides a comprehensive evaluation of the seasonal prediction skill of CFS/CFSv2 for DGPI in the WNP, highlighting its limitations, particularly in recent operational forecasts.
• Demonstrates the significant potential of deep learning, specifically CNNs combined with transfer learning from CMIP6 and reanalysis data, to substantially enhance seasonal forecasting of tropical cyclone genesis potential.
• Offers a novel and effective approach to improve the predictive skill of large-scale environmental conditions crucial for tropical cyclone activity, outperforming state-of-the-art operational models.

Funding

• Not specified in the provided text.

Citation

@article{Peng2025Assessing,
  author = {Peng, Shanchi and Wang, Chao and Wu, Liguang and Zhao, Haikun and Cao, Jian},
  title = {Assessing seasonal prediction of DGPI over the Western North Pacific by the Climate Forecast System and its improvement using deep learning},
  journal = {Atmospheric Research},
  year = {2025},
  doi = {10.1016/j.atmosres.2025.108628},
  url = {https://doi.org/10.1016/j.atmosres.2025.108628}
}