*Daisuke Matsuoka1,2, Masuo Nakano1, Daisuke Sugiyama1, Seiichi Uchida3
(1.Japan Agency for Marine-Earth Science and Technology, 2.Japan Science and Technoloy Agency, 3.Kyushu University)
Keywords:Deep learning, Tropical cyclogenesis, Prediction of occurrence, cloud resolving global atmospheric model
In recent years, deep learning, one of the machine learning methods based on neural networks, has been applying to image pattern recognition. In the present study, we investigate the probability of predicting Tropical Cyclones (TCs) 14 days prior from long-term global atmospheric simulation data (only Outgoing Longwave Radiation) using deep convolutional neural networks (CNNs). Our deep CNNs train 50,000 TC data including its precursor and 500,000 not TC data (center of low pressure) generated by TC tracking algorithm. As a result, we succeeded in predicting the precursors of TCs seven and 14 days before their formation with a Recall of 92.0%. Although seasonal and spatial predictability of precursor of TCs are strongly correlated with the number of training data, in some seas and/or seasons, high accuracy is obtained despite the small amount of training data.