[A-TT30] Machine Learning Techniques in Weather, Climate, Ocean, Hydrology and Disease Predictions

[ATT30-P01] 変分オートエンコーダによる急潮の分類

*青木 邦弘¹、中野 英之¹、広瀬 成章¹、坂本 圭¹、碓氷 典久¹、豊田 隆寛¹、浦川 昇吾¹、川上 雄真¹ (1.気象庁　気象研究所)

キーワード：クラスタリング、深層生成モデル、急潮、黒潮

Kyucho is well known as a coastal oceanic abrupt event causing a strong surface current that often destroys fishing set-nets. The prediction of Kyucho requires specifying its precursor, but, in the case where Kyucho has multiple flow patterns, their triggers can be different from each other, and, this applies to Muroto, a southern coastal area of Shikoku, Japan. However, since the flow patterns of Muroto-Kyucho have been classified by "human eyes" from limited oceanic data, the whole picture of the rigorous flow patterns has still been unclear. In this study, we identify the flow patterns accompanied by Muroto-Kyucho in a statistical manner, leveraging the variational autoencoder, a deep generative model. The method can reconstruct the spatial structure of the oceanic flow from the so-called latent space which stores compressed information of the high dimensional oceanic flow data. We apply the method to the surface current data in 1960-2007 of FORA-60JPN, a reanalysis dataset supplied by Meteorological Research Institute, Japan Meteorological Agency. The latent space shows at least three modes, whose corresponding reconstructions indicate the different flow patterns that include the strong coastal current expressing Muroto-Kyucho: 1) Coastal jet like structure successive from Kii peninsula to Muroto, 2) Cyclonic eddy-like structure just southeast of Muroto, 3) Intrusion of Kuroshio current from southwest of Muroto. Pattern-1 is consistent with the typical structure in the coastal surface density flow or coastal trapped waves, but Pattern-2 and 3 suggest a presence of different dynamics. We will discuss potential precursors of each classified Muroto-Kyucho at the meeting.