Predicting Tropical Cyclone Intensity Using a Convolutional Neural Network and 20 Years of IMERG Satellite Rainfall Data

Abstract The prediction of tropical cyclone (TC) intensity change remains one of the greatest challenges for forecasters. The Statistical Hurricane Intensity Prediction Scheme (SHIPS) is one of the most accurate models used in operational centers. The current version of SHIPS uses predictors including climatology and persistence, environmental conditions, and infrared satellite information. One critical piece of information that is largely missing from SHIPS is the rainfall and structural features of TCs. In this study, a novel Hurricane Convolutional Neural Network (HCNN) model is proposed to predict future TC intensity by using satellite rainfall images and existing SHIPS predictors. A 20-year (2000–2019) satellite rainfall dataset is obtained from the NASA Integrated Multi-satellite Retrievals for the Global Precipitation Measurement (GPM) mission (IMERG) product for TCs from the Atlantic basin. The HCNN model is tested for 3 different radii of the IMERG data from the TC center and 200-km is selected. The model is trained using satellite images for TCs from 2000–2017 and tested using TCs from 2018–2019. Relative to a multiple linear regression model with SHIPS predictors trained using the same training sample and tested using the same test sample as used for the HCNN, the HCNN model with satellite rainfall input significantly improves forecasts by 9–13%, 8–18%, and 5–9% for all TCs, major hurricanes, and intensifying TCs, respectively, at 6–24 hour forecast intervals. Further experiments show that the HCNN can better utilize rainfall structural information than a multi-linear regression model integrating SHIPS and rainfall predictors.