17:15 〜 18:45
[MGI24-P06] Predictability of Kuroshio path using deep learning based uNet model at 30- and 60- days lead time
キーワード：Kuroshio path prediction, Deep learning, uNet
Recent deep learning methods have significantly outperformed in many geoscience-related applications. In this study, we assess the predictability of the Kuroshio path over 22.7°–40.5°N and 120°–150°E using deep learning-based encoder-decoder models: “uNet” at 30- and 60-day lead times. The proposed uNet model is trained using sea surface height (SSH) and vertically averaged (40 m–110 m depth) subsurface temperature (SubST) data of the ocean reanalysis from 2019 to 2022 and tested on 2023 (January to September). Various versions of “uNet” were experimented with, varying in the activation, loss function, convolution modules, and input attributes. The results from the best models were analyzed using the 2-D correlation coefficient (CC), analogous to the image similarity coefficient calculated between the target and the predicted Kuroshio path over an ensemble of the models. The results from the study show high CC values for the upstream side, moderate values for the meandering region, and moderate to poor values for the northward extension. Even though the proposed “uNet” model exhibits moderate to poor CC in the meandering and northward regions, in a few cases the predicted path matches very well with the target Kuroshio path. These results are comparable with recent deep learning-based studies. Further experimentation using more training samples, loss functions, and temporal convolutional modules suggests promising potential in performance improvements.