Exploring Sub-Basin Seasonal North Atlantic Tropical Cyclone Projections Using Machine Learning

Both recent and historical work has explored the regulation of seasonal tropical cyclone (TC) activity by the El Niño-Southern Oscillation (ENSO), the Atlantic Meridional Mode (AMM), and the Atlantic Multidecadal Variability (AMV) teleconnections. Further, recent work has begun exploring the utility of the sea surface temperature differences between the Atlantic main development region (MDR) and equatorial Pacific in predicting North Atlantic seasonal TC activity, with some success on a basin level. These studies have motivated the development of a sub-basin level projection of TC activity using linear techniques that has shown promising results, with R² values as high as 0.49 for some defined sub-basins.

Despite these advances, it is likely that many of the relationships among these SST structures maintain some fraction of nonlinear variability explained, such that linear techniques may not be capturing the full potential of SST indices in predicting sub-basin North Atlantic TC activity. Thus, in this study we will evaluate the utility of machine learning techniques, namely support vector machines and random forests, in predicting seasonal sub-basin level TC activity for the North Atlantic. Following prior studies, we have split the North Atlantic basin into five sub-regions, the Gulf of Mexico (GOM), Caribbean (CAR), eastern United States (EUS), subtropical North Atlantic (SNA), and the tropical North Atlantic (TNA). SST-derived predictors were computed from the Hadley Centre Global Sea Ice and Sea Surface Temperature (v1) reanalysis from 1966-2023 that characterize SST structures from the MDR, Niño 3, and Atlantic 3 regions, as well as an AMM index. These predictors were used to forecast accumulated cyclone energy (ACE) levels for a given upcoming TC season, categorizing the season as below average, average, or above average in its activity level. The resulting analyses were compared against existing linear baselines to quantify the benefits of the machine learning techniques in this application.