Diagnosing Tropical Cyclone Rapid Intensification Using Kernel Methods and Reanalysis Datasets

Tropical cyclone rapid intensification (RI) continues to be a problem that eludes operational forecasters. Recent work in this area has revealed the value of applying machine learning techniques to classifying storms as RI or non-RI at 24-hours lead time. However, that work showed that differing reanalysis datasets represented the storms in unique ways, offering different discrimination capability and unique predictor sets that are important for RI. The scope of this research is to identify factors important for RI that are consistent among three reanalysis datasets, as these are likely the fields that will provide the greatest discrimination capability. An S-mode rotated principal component analysis was used to formulate unique patterns within RI and non-RI storms, and the resulting RPC scores were used to train a support vector machine classification algorithm that yielded binary RI occurrence output. Base-state meteorological variables (geopotential height, temperature, u and v wind components, vertical velocity, and relative humidity) at single horizontal levels were tested individually as predictors for the SVM. Base-state fields that were consistently good at discriminating RI events from non-RI events among all three reanalysis datasets were deemed most useful for RI classification and will be considered for future forecast applications