5:15 PM - 7:15 PM
[ATT35-P05] Machine Learning-Based Bias Correction to Improve Marine Heatwave Forecasts in the Korean Marginal Sea
Keywords:Marine heatwave, Machine learning, Artificial intelligence, LSTM
Global warming has intensified marine heatwaves, significantly affecting ecosystems and economies. In South Korea, marine heatwaves have occurred almost annually since the 2010s, with the extreme event in 2018 causing damages of approximately 60 billion KRW. In response, the Korea Institute of Ocean Science and Technology (KIOST), the Korea Polar Research Institute (KOPRI), Pukyong National University, and other collaborating institutes have developed the Korean Marine Heatwave Prediction System (KMHPS) based on an air-sea coupled model. While KMHPS effectively captures the onset and decay of marine heatwaves, it underestimates sea surface temperature (SST).
This study hypothesizes that the underestimation to the coarse vertical resolution of the upper ocean layers in KMHPS. Although dynamical models require substantial computational resources, machine learning provides a more efficient alternative. To address the underestimation, a long short-term memory (LSTM) model was developed using upper-layer SST data (0–5 m) from the GLORYS reanalysis dataset, which offers a finer vertical resolution (~1 m) than KMHPS at the surface(5 m). Additionally, shortwave radiation, a key driver of marine heatwaves, was incorporated to improve predictive accuracy.
Evaluation from July to August 2023 demonstrated limited improvements in the East/Japan Sea and Yellow Sea but yielded accuracy enhancements during the marine heatwave events. The south of the Korea peninsula exhibited the most significant improvements, reducing SST errors by over 1°C. Furthermore, the correction improved over time, leading to a 0.7°C reduction in root-mean-square error (RMSE) after 14 days compared to KMHPS. Incorporating shortwave radiation further improved prediction accuracy. This suggests that integrating multiple influential factors is expected to enhance bias correction and improve marine heatwave forecasting.
This study hypothesizes that the underestimation to the coarse vertical resolution of the upper ocean layers in KMHPS. Although dynamical models require substantial computational resources, machine learning provides a more efficient alternative. To address the underestimation, a long short-term memory (LSTM) model was developed using upper-layer SST data (0–5 m) from the GLORYS reanalysis dataset, which offers a finer vertical resolution (~1 m) than KMHPS at the surface(5 m). Additionally, shortwave radiation, a key driver of marine heatwaves, was incorporated to improve predictive accuracy.
Evaluation from July to August 2023 demonstrated limited improvements in the East/Japan Sea and Yellow Sea but yielded accuracy enhancements during the marine heatwave events. The south of the Korea peninsula exhibited the most significant improvements, reducing SST errors by over 1°C. Furthermore, the correction improved over time, leading to a 0.7°C reduction in root-mean-square error (RMSE) after 14 days compared to KMHPS. Incorporating shortwave radiation further improved prediction accuracy. This suggests that integrating multiple influential factors is expected to enhance bias correction and improve marine heatwave forecasting.