An Updated Atlantic Basin Tropical Cyclone Rapid Intensification Scheme Using Machine Learning and Operational Forecast Data

Recent work has quantified the ability of machine learning methods to identify the onset of rapid intensification (RI) of Atlantic tropical cyclones (TCs) at 24-hours lead time. Here we provide an updated algorithm, where Global Forecast System analysis fields are blended with predictors from the Statistical Hurricane Intensity Prediction Scheme Rapid Intensification Index (SHIPS-RII) to develop a machine learning ensemble for Atlantic RI prediction. Kernel principal component (KPC) analysis was used on TC- centric GFS analyses to identify KPC scores that yield the greatest separation between RI and non-RI storms. These KPC scores were combined with the full set of SHIPS-RII predictors (roughly 120 initial predictors) and run through a forward feature selection to identify the optimal combination of predictors for RI prediction. After feature selection, predictors were used input to a variety of machine learning method configurations to identify the combinations that provided the best forecast skill (based on the Brier Skill Score – BSS). Machine learning configurations that yielded the largest BSS were retained as part of a machine learning ensemble (9 members were retained). The KPCA yielded optimal separation when combined with support vector machine classifiers with varying cost and separation values in the radial basis function kernel. Results were promising in the training phase, as Brier skill scores exceeded 0.46 in the training phase (relative to a baseline 0.25 BSS in the SHIPS-RII operational ensemble currently utilized by the National Hurricane Center). Updated results are presented for the operational and 8 experimental RI definitions.