Estimating Tropical Cyclone Intensity in Passive Microwave Images Using Deep Learning Models

Accurately estimating tropical cyclone intensities in real time is a consistent forecasting challenge, despite improvements over the last 40 years. Common techniques suffer from human error, time constraints, and complexity. Ideally, challenges could be mitigated through automated, timely, and generalizable processes that reduce root mean squared errors. This study evaluates one such technique, applying deep learning models to passive microwave images. Specifically, deep convolutional neural networks estimated cyclone intensities of Category 1 or below, as previous studies showed that models faltered with infrared images categorized by wind speeds less than 80 kt. Thus, to determine whether the better defined cyclonic structures in passive microwave images could improve accuracy and error rates, this study examined interpolated 85 GHz passive microwave brightness temperatures from the Level 1C GPM Common Calibrated Brightness Temperatures dataset (1997-2016; NASA’s Earth Observing Systems). Two different architectures were employed, one that was a traditional convolutional neural network model and predicted probabilities for each represented intensity class. A second convolutional model relied on linear regression to predict a single intensity for every test image and was measured by root mean squared error. Additionally, image processing was tested using 1- and 5 kt bin classifications as well as nearest neighbor and bilinear interpolations. Preliminary studies into Northern Atlantic and Pacific Ocean observations found that the traditional convolutional neural network models achieved a 22% accuracy rate at best, while the regression models achieved root mean squared errors <14 kt, suggesting that predictions based on linear regressions are recommended for intensity estimation among passive microwave images. In particular, models analyzed in this study performed best when making intensity estimations of tropical cyclones whose lifetime maximum intensities were below 80-kt and achieved the most favorable results when images were processed using bilinear interpolation in place of nearest neighbor interpolation. Future investigations into these results will include masking brightness temperatures above freezing, including all wind speed classes to increase the number of image inputs from storms with high lifetime maximum intensities, and training models to better recognize unique storm features that distinguish between poorly and well-organized systems.