[A-AS02] Advances in Tropical Cyclone Research: Past, Present, and Future

[AAS02-07] Statistical prediction of tropical cyclone rapid intensification with an explainable AI

★Invited Papers

*Takeshi Horinouchi¹, Takashi Yanase², Yuiko Ohta², Daisuke Matsuoka⁵, Asanobu Kitamoto⁴, Udai Shimada⁶, Ryuji Yoshida³, Fudeyasu Hironori³ (1.Faculty of Environmental Earth Science, Hokkaido University, 2.Artificial Intelligence Laboratory, Fujitsu Research, Fujitsu Limited, 3.TRC, Yokohama National University, 4.National Institute of Informatics, 5.JAMSTEC, 6.MRI, JMA)

Keywords:Typhoon, Rapid intensification, Explainable AI

Imperfectness of the state-of-the-art intensity forecasting of tropical cyclones (TCs) necessitates independent rapid intensification (RI) prediction schemes. Here we report one derived with an explainable artificial intelligence Wide Learning (WL). The scheme, named Wide Learning-based TC Rapid intensification Prediction Scheme (WRPS) Version 1 (WRPS1), predicts RI in the Western North Pacific by using twelve predictor variables representing environmental conditions and the state of TCs. Its prediction is based on a score that is a linear combination of whether or not (1 or 0) joint conditions on ranges of multiple variables are met, which is reproducible without WL. Relying on joint conditions allows WRPS to handle nonlinearity and inter-dependence among predictors, and the simpleness of the conditions provides explainability. A method to map an RI-prediction score to its probability is proposed and is used in WRPS. It is suggested that handling predictors favorable to RI when having moderate values, such as the current intensity, is a key for good RI prediction. It is demonstrated that quantifying the contribution of each predictor to the WRPS score helps one elucidate how the predictors jointly facilitated or hindered RI for each prediction case. The performance of WRPS1 is compared with RI predictions using the linear discriminant analysis, and WRPS1 is shown to perform well without using track predictions. The multiple linear regression analysis, which is customarily used for intensity prediction but not for RI prediction, is shown to perform well if the fraction of RI cases is increased when conducting regression.