9:30 AM - 9:45 AM
[MGI27-03] Deep learning approach to subseasonal prediction of the western North Pacific subtropical high: transfer and multitask learning
Keywords:western North Pacific subtropical high, Deep Learning, Transfer Learning, Multitask Learning
During the boreal summer, the western North Pacific subtropical high (WNPSH) is prominent in the Northwest Pacific, significantly influencing heatwaves, typhoon tracks, and the Baiu front. Accurate prediction of the WNPSH and understanding its driving mechanisms are crucial for advancing our knowledge of the Asian summer monsoon system. The WNPSH exhibits substantial variability over the region south of Japan, with time scales ranging from daily to interannual fluctuations. This variability is complex, driven by interactions between tropical and mid-latitude systems, posing challenges for numerical model-based predictions. In this study, we construct a data-driven approach utilizing deep learning techniques, specifically transfer learning and multitask learning, to improve subseasonal predictions of the WNPSH with a lead time of approximately one month. To capture diverse representations, we employed transfer learning by pretraining on a large-scale ensemble dataset (d4PDF) spanning thousands of years, followed by fine-tuning using ERA5 reanalysis data. A supervised learning framework based on convolutional neural networks (CNNs) was adopted, incorporating multitask learning to simultaneously predict the WNPSH and related phenomena such as the Boreal Summer Intraseasonal Oscillation (BSISO). This multitask approach yielded higher predictive skills compared to models trained solely on ERA5 data. Furthermore, analyzing task-to-task skill relationships revealed that the predictability of the WNPSH is influenced by factors such as the BSISO phases. We will explore relationships between other teleconnection patterns to further elucidate predictive factors.