Using Deep Learning to Estimate Tropical Cyclone Intensity from 89-GHz Band Microwave Satellite Imagery

Tropical cyclone (TC) forecasting and analysis, as a field of research, is still looking for ways to objectively integrate the rich information of multichannel microwave, infrared and visible frequency satellite sensors. Among these, the 85-92 GHz band, which depicts major TC structures such as eye walls and rain bands, is a critical tool for subjective analysis but is not used by forecasters or analysts in the official, analytical approach for estimating TC intensity.

In this study, a Deep Learning Convolutional Neural Network model is used to explore the possibilities for estimating TC intensity from images in the 85-92 GHz band. The model, called “DeepMicroNet,” also has unique properties such as a probabilistic output, the ability to operate from partial scans (including bad scan lines) and resiliency to inaccurate TC center-fixes. Results suggest that the 85-92 GHz band has value for estimating TC intensity with better precision than was previously known, especially for Category 3-5 TCs. Overall the model has an accuracy that approaches existing methods to estimate TC intensity from satellites. Robust model training with a global Best Track data set (the official record of TC history) and subsequent testing with aircraft reconnaissance data yields results precise enough to demonstrate a bias in the Best Track record itself (an underestimation of intensity) in the intensity range of 35-45 kt. The results here show the tremendous promise of Deep Learning applied to TC intensity analysis to produce superior forms of operational information, with clear pathways for application to other frequencies as well.