Probabilistic Rapid Intensification Prediction with Convolutional Neural Networks and HWRF

Rapid intensification of tropical cyclones is favored with certain spatial signatures in the wind and thermodynamic fields of the storm. Increased symmetry in these fields often indicates increased storm intensity, but in earlier stages of rapid intensification, particularly in an environment containing vertical wind shear, storm structures are more complex and therefore offer various morphological pathways toward intensification. Current statistical rapid intensification models utilize spatial averages of certain features (e.g., environmental vertical wind shear over some radius) to diagnose these spatial patterns and have shown skill in discriminating rapid intensification events. Further skill increases may be realized by using deep learning models to identify more complex spatial patterns directly from the gridded storm fields. In this project, we have collected Hurricane Weather Research and Forecasting (HWRF) reforecasts with the 2018 operational configuration of HWRF over the hurricane seasons from 2015-2017 in the Atlantic and Eastern Pacific basins. Given the equivalent potential temperature, radial wind, and tangential wind fields from HWRF at the 850-, 700-, and 500-hPa levels, the convolutional neural network predicts the distribution of possible changes in maximum wind speed over a 24-h period. We test both a basic convolutional neural network and a ResNet-style architecture. They are compared with neural network and random forest models using researcher-selected features from HWRF as input. Real-time HWRF data from the 2018 and 2019 hurricane seasons are used to test each of the deep learning models and compare them with the HWRF and other operational intensity forecasts. In addition to evaluating the performance of the models, we also interpret the convolutional neural networks and traditional machine learning models to identify which spatial features are associated with rapid intensification. We use saliency maps, permutation variable importance, and partial dependence plots for the evaluation of important features in the neural networks.