AbstractAbstractTropical cyclones (TCs) are an extreme weather hazard that can have major impacts on coastal and inland populations. Surveillance by meteorological satellites has essentially mitigated the problem of TC detection; however, estimating TC intensity via remote sensing techniques is more challenging. Exploring different sensor frequencies and developing novel ways to interrogate that data are essential steps toward improving the intensity analysis of TCs. This study employs a convolutional neural network (CNN) machine learning method trained on previously unexplored 183-GHz microwave imagery to estimate TC intensity. A CNN can identify two-dimensional features in satellite imagery that are indicative of a TC’s current intensity. The novel approach, coined “D-MINT183,” is similar to an earlier version [Deep Multispectral Intensity of TCs estimator (D-MINT)] in that it uses infrared imagery and environmental scalar predictors, but D-MINT183 uses 183-GHz microwave imagery, while D-MINT uses legacy 37 and 89 GHz. The D-MINT183 model intensity estimates are calculated for different types of low-Earth-orbiting (LEO) meteorological satellites: conical scanning and cross-track scanning, and the experimental TROPICS CubeSats. Results show that D-MINT183 TC intensity estimates from conical scanning satellites are generally as skillful as D-MINT. For cross-track scanning satellites and TROPICS, when D-MINT is not available, TC intensity estimates from D-MINT183 are more skillful than Deep infrared (IR) Intensity of TCs estimator (D-PRINT), a CNN which does not use microwave imagery. Thus, D-MINT183 can increase the frequency of more accurate, microwave-based estimates of TC intensity. This work also demonstrates the importance of explainable artificial intelligence, with examples highlighting the situationally dependent impacts of 183-GHz imagery on the D-MINT183 intensity estimates.Significance StatementThis study develops a machine learning method to estimate tropical cyclone (TC) intensity using novel satellite observations. The resultant model uses a convolutional neural network (CNN) which can identify two-dimensional features in 183-GHz microwave imagery along with geostationary infrared imagery that are indicative of the current TC intensity. Model performance shows that in several TC basins, the inclusion of the previously unexplored (but operationally available) 183-GHz water vapor channels on current and future low-Earth-orbiting (LEO) satellites can provide information that can complement alternative satellite-based estimates of TC intensity. This is significant since G-band frequencies can provide suitable spatial resolutions with much smaller antenna size requirements and lower costs, meaning the data are likely to be more common on emerging and planned LEO CubeSat constellations.
