A deep learning ensemble approach for predicting tropical cyclone rapid intensification

Abstract

Predicting rapid intensification (RI) of tropical cyclones (TCs) is critical in operational forecasting. Statistical schemes rely on human-driven feature extraction and predictor correlation to predict TC intensities. Deep learning provides an opportunity to further improve the prediction if data, including satellite images of TC convection and conventional environmental predictors, can be properly integrated by deep neural networks. This study shows that deep learning yields enhanced intensity and RI prediction performance by simultaneously handling the human-defined environmental/TC-related parameters and information extracted from satellite images. From operational and practical perspectives, we use an ensemble of 20 deep-learning models with different neural network designs and input combinations to predict intensity distributions at +24 h. With the intensity distribution based on the ensemble forecast, forecasters can easily predict a deterministic intensity value demanded in operations and be aware of the chance of RI and the prediction uncertainty. Compared with the operational forecasts provided for western Pacific TCs, the results of the deep learning ensemble achieve higher RI detection probabilities and lower false-alarm rates.