A microwave-based machine learning approach for predicting eyewall replacement cycles

Abstract

Eyewall replacement cycles (ERCs) greatly increase the destructive potential of tropical cyclones (TCs) by affecting the maximum wind speed, wind field size, and storm surge severity while simultaneously reducing confidence in TC forecasts, most prominently in intensity forecasting. Machine learning (ML) presents new opportunities to improve current forecasting and predictive capabilities, and its application will benefit forecasters and ultimately the public. The objective of this project was to create a proof-of-concept ML convolutional neural network (CNN) to predict ERCs using the 89 GHz microwave band for training and testing. The training set was comprised of North Atlantic basin (NATL) storms from 1999 to 2009. The testing set included NATL storms from 2019 to 2022. Twelve models were created, together known as the CNN Ensemble for Predicting Eyewall Replacement Cycles (CE-PERCY), with each individual member achieving at least 80 % in-training accuracy. Two versions were created: versions A and B. Using synthetic aperture radar, land-based radar, aircraft reconnaissance, Microwave-based Probability of ERC (M-PERC), National Hurricane Center reports, and microwave imagery, ERC analysis was conducted on the testing set. 28 ERCs were identified throughout 14 hurricanes from 2019 to 2022. CE-PERCY performs well for a proof-of-concept, with versions A and B predicting 21 and 23 ERCs, respectively. This project successfully introduces a foundation for using ML CNNs in ERC prediction, demonstrates the viability of the technique, and proves that a large enough dataset of microwave imagery can be used in this specific application.