Understanding Tropical Cyclone Transitions using JPSS Data and Deep Learning

The transition between tropical cyclone types has associated changes in storm structure, and so understanding and predicting these transitions is crucial for improving tropical cyclone (TC) forecasts and mitigating their impacts. For example, the transition from a tropical to extratropical (ET) cyclone often involves increased rainfall and an expansion of the wind field. A notable example is Hurricane Sandy (2012), which, during its ET transition, approached the U.S. East Coast with a very large wind field, leading to a significant storm surge even though it was only a category 1 hurricane just before the ET transition. Given the potential impacts, official forecasts from the National Hurricane Center (NHC), Central Pacific Hurricane Center (CPHC), and Joint Typhoon Warning Center (JTWC) include storm type classifications to better understand these transitions.

Global weather forecast models do not always correctly diagnose storm transitions. Thus, given that the majority of global TCs occur in remote areas, satellite data are an important additional source of information for diagnosing TC transitions. Microwave sounders on low-Earth-orbiting (LEO) satellites, such as the Advanced Technology Microwave Sounder (ATMS) on JPSS satellites and the Advanced Microwave Sounding Unit (AMSU) on NOAA and MetOp satellites, can penetrate clouds and provide crucial information on TC intensity and structure that is not available from visible and infrared imagery. Despite the higher latencies of polar-orbiting data (typically 1-2 hours) compared to geostationary satellite data (10-15 minutes), the increasing availability of microwave sounder data on legacy satellites and smallsats has opened new avenues for employing machine learning techniques to classify TCs, potentially enhancing operational storm type forecasts.

In this presentation, we will discuss a new storm type classification machine learning model, the JPSS Extratropical Transition Classification (JETClass). The model is built on the operational ETClass method, which utilizes Linear Discriminant Analysis with predictors from infrared imagery and NCEP Global Forecast System (GFS) model, and is part of the NHC’s Statistical Hurricane Intensity Prediction Scheme (SHIPS). JETClass leverages neural network architectures to integrate predictors from ATMS/AMSU data, Automated Tropical Cyclone Forecast (ATCF) system best track data, GFS model fields, and geostationary satellite imagery. The model is trained to classify storms into three categories: tropical, subtropical, and extratropical. Evaluation of the model using the Peirce Skill Score shows a promising score of 0.77 on independent cases, indicating strong potential for operational use. In addition, the high model skill indicates that JETClass can be used to provide insight into the physical processes associated with the TC transitions. Explainable Artificial Intelligence (AI) techniques will be used to determine the factors most important for the transition between TC types. Furthermore, we will discuss the model's performance during the 2024 storm season as a near real-time product.

Disclaimer: The scientific results and conclusions, as well as any views or opinions expressed herein, are those of the author(s) and do not necessarily reflect those of NOAA or the Department of Commerce.