Application of a Deep Learning Neural Network to Remotely-Sensed TC Intensity

Deep learning neural network software has made enormous strides in commercial applications, from web searching, to billion dollar hedge funds, to self-driving cars. However, the applications for meteorological research are just beginning. Here we explore the capability of a deep learning image classification scheme to estimate TC parameters such as maximum sustained wind and minimum sea level pressure using only microwave and infrared satellite imagery, trained on best track intensities from NHC and JTWC. Algorithm performance is compared to alternative methods such as CIMSS SATCON and the Dvorak method. In addition, the versatility of the deep learning approach can also demonstrate inherent sensitivities of the satellite channels, such as a relatively high precision of the 85-GHz channel in the 65-85 knot intensity range and a lower precision in the 30-50 knot range. This unique angle of investigation can be used to confirm subjective understandings of satellite images and also reveal areas of untapped potential.