Kim et al. (2025) Deep learning-based prediction of cold surge frequency over South Korea

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-12-03
• Authors: Eung‐Sup Kim, Joonlee Lee, Jina Hur, Sera Jo, Yong-Seok Kim, Kyo‐Moon Shim, Joong‐Bae Ahn
• DOI: 10.1038/s41598-025-28608-z

Research Groups

• Climate Change Division, National Institute of Agricultural Sciences, Wanju, South Korea
• Department of Civil, Urban, Earth, and Environmental Engineering, Ulsan National Institute of Science and Technology, Ulsan, South Korea
• Department of Atmospheric Sciences, Pusan National University, Busan, South Korea

Short Summary

This study develops a hybrid deep learning framework combining a coupled general circulation model with a Long Short-Term Memory neural network to improve seasonal prediction of winter cold surge frequency over South Korea, demonstrating significantly enhanced prediction skill and revealing a temporal shift in dominant teleconnection drivers.

Objective

• To enhance the seasonal prediction skill of winter Cold Surge Days (CSDs) in South Korea by combining a dynamical model (Coupled General Circulation Model - CGCM) with a deep learning model (Long Short-Term Memory - LSTM), demonstrating the complementarity between physical simulation and data-driven learning.
• To examine changes in prediction performance and explore their possible connections with shifts in the dominant teleconnection drivers, focusing on the relative roles of oceanic and atmospheric patterns.

Study Configuration

• Spatial Scale: South Korea (region 125–130°E and 35–40°N), with implications for broader East Asia.
• Temporal Scale: Winter (December–February) seasonal predictions, covering 42 winters from 1980/81 to 2021/22.

Methodology and Data

• Models used:
  • Long Short-Term Memory (LSTM) neural network.
  • Pusan National University/Rural Development Administration (PNU/RDA) Coupled General Circulation Model (CGCM), comprising:
    • Atmosphere: Community Climate Model (CCM3, spectral truncation T42, 18 hybrid sigma-pressure levels, top: 2.917 hPa).
    • Land: Land Surface Model (LSM, 6 levels).
    • Ocean: Modular Ocean Model (MOM3, horizontal longitude: 2.8125°, low latitude: ~0.7°, mid latitude: ~1.4°, high latitude: ~2.8°, 40 levels, top: 10 m, bottom: 5258 m).
    • Sea-Ice: Elastic-Viscous-Plastic Model (EVP, 3 levels).
  • SHapley Additive exPlanations (SHAP) for model interpretability.
  • Multiple Linear Regression (MLR) for comparison.
• Data sources:
  • CGCM-predicted daily mean 2-meter temperature (10 ensemble members), bias-corrected using ERA5 reanalysis.
  • 24 selected climate indices (from an initial 30 candidates) covering lead times from March to October, including Trans-Niño Index (TNI), El Niño Modoki Index (EMI), Niño 1.2 index, Pacific–Japan (PJ) index, East Atlantic–Western Russia (EAWR) index, Circumglobal Teleconnection (CGT) index, Scandinavia pattern (SCAND), Western Pacific pattern (WP), Southern Oscillation Index (SOI), North Pacific (NP), and Southern Annular Mode Index (SAMI).
  • Observed daily temperature data from 56 Automated Synoptic Observing System (ASOS) stations operated by the Korea Meteorological Administration (KMA).
  • ERA5 reanalysis (European Centre for Medium-Range Weather Forecasts) at 0.1° horizontal resolution.
  • Cold Surge Day (CSD) definition: daily mean temperature decrease exceeding 1.5 standard deviations (σ) within two days, with termination when temperature returns above –0.5σ.
• Validation: Leave-One-Year-Out Cross-Validation (LOYOCV) and an independent train–test split (1980–2001 for training, 2002–2021 for validation).

Main Results

• The hybrid LSTM framework significantly improved seasonal prediction skill for winter CSDs over South Korea compared to the PNU/RDA CGCM alone. Over the entire period (1980–2021), LSTM achieved a Pearson correlation coefficient (CORR) of 0.62 (statistically significant at 99% confidence level) and a root mean square error (RMSE) of 6.95, compared to the CGCM's CORR of 0.33 and RMSE of 8.33.
• Prediction skill was notably higher in the recent period (P2: 2001–2021), with LSTM's CORR increasing from 0.27 in P1 (1980–2000) to 0.88 in P2, and normalized RMSE (nRMSE) decreasing from 1.18 to 0.51.
• SHAP analysis identified the Scandinavia pattern (SCAND10), Western Pacific pattern (WP7), and Southern Oscillation Index (SOI_9) as consistently strong influences on CSD prediction.
• A temporal shift in dominant teleconnection drivers was observed, with atmospheric indices showing increased contributions to prediction skill in P2 compared to oceanic indices, consistent with the influence of Arctic amplification and jet stream changes.
• Sensitivity experiments confirmed that combining CGCM predictions with climate indices (EXP1) yielded the best performance, outperforming models using only climate indices (EXP2) or only CGCM predictions (EXP3), highlighting the complementary value of both approaches.

Contributions

• Developed a novel hybrid prediction framework combining a dynamical Coupled General Circulation Model (CGCM) with a deep learning Long Short-Term Memory (LSTM) model for seasonal prediction of cold surge frequency.
• Demonstrated substantial and robust improvements in cold surge prediction skill over South Korea, particularly in recent decades, compared to traditional dynamical models.
• Provided interpretability of the prediction model using SHAP analysis, identifying key teleconnection patterns influencing cold surges.
• Revealed a temporal shift in the dominant drivers of cold surge predictability from oceanic to atmospheric patterns, linking it to evolving large-scale climate dynamics like Arctic amplification.
• Highlighted the complementary value of integrating physical simulations and data-driven learning for climate risk management and early prediction of extreme events in East Asia.

Funding

• Cooperative Research Program for Agriculture Science and Technology Development (Project No. PJ01746002), Rural Development Administration, Republic of Korea.
• 2025 RDA Fellowship Program of National Institute of Agricultural Sciences, Rural Development Administration, Republic of Korea.

Citation

@article{Kim2025Deep,
  author = {Kim, Eung‐Sup and Lee, Joonlee and Hur, Jina and Jo, Sera and Kim, Yong-Seok and Shim, Kyo‐Moon and Ahn, Joong‐Bae},
  title = {Deep learning-based prediction of cold surge frequency over South Korea},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-28608-z},
  url = {https://doi.org/10.1038/s41598-025-28608-z}
}