[M-AG39] 海洋地球インフォマティクス

[MAG39-03] Automatic detection of weather front around Japan using deep convolutional neural network

*松岡 大祐^1,2、杉本 志織 ¹、中川 友進¹、荒木 文明¹、川原 慎太郎¹、尾上 洋介⁴、飯山 将晃³、小山田 耕二³ (1.海洋研究開発機構、2.科学技術振興機構、3.日本大学、4.京都大学)

キーワード：ディープラーニング、畳み込みニューラルネットワーク、気象前線抽出

In this study, we automatically detect stationary front from weather forecast simulation data (GPV/MSM) using U-Net deep convolutional neural network. Our U-Net trains ten years weather data (precipitation, sea level pressure, relative humidity, air temperature, and wind velocity) and the position of weather front as the ground truth. As a result of applying the trained U-Net to the untrained one year data, our approach succeeded in accurately detecting Baiu front and autumn rain front except when Typhoon occurred. Moreover, wind velocity (zonal and meridional component) and relative humidity at 1000 hPa play an important role to obtain high detection performance. Our approach is also able to apply to weather simulation data which the weather front is not associated with.