Predicting Tropical Cyclone Rapid Intensification From Satellite Microwave Data and Neural Networks

A new method to analyze the potential for rapid intensity change in tropical cyclones (TC) is presented. The method is based on satellite observations of precipitation derived from microwave (MW) radiometers. The approach is intended to condense the information in the environment and in the vortex using a low wavenumber representation of the rain index (RaIn, a multichannel nonlinear combination of passive MW observations), and train a deep-learning, multilayer neural network (NN) with the RaIn and the changes in the wind over the next 24 h. The resulting NN exhibits a near-perfect ability to identify rapid intensification (RI: changes in the hurricane wind speed in excess of 30 knots within a 24-h period). It is found that the spatial structure and amounts of the columnar water condensate within the extended environment is necessary to capture the most important information regarding the RI process. Analyses of the NN structure provide new insight into the physics of TC and can help improve model forecasting. Environmental conditions as far as 1050 km from the TC center might affect the process of RI by at least three physical processes: absolute angular momentum inflow, wind shear stabilization, and steering the outflow jets in the upper troposphere. The findings can be used to build a RI discriminant (RID) for real-time operations.