A Spatially-Aware North Atlantic Tropical Cyclogenesis Index Using Machine Learning

A tropical cyclone genesis index (TCGI) is a tool used to evaluate the potential of tropical cyclone (TC) formation based on large-scale environmental conditions at varying timescales. TCGIs can be applied to reanalysis products to assess genesis climatology and variability, and they can also be applied to climate models to estimate TC frequency in a future climate. Past TCGIs use variants of thermodynamic- and dynamic-based environmental predictors such as vertical wind shear, vorticity, moisture, and potential intensity. However, these predictors can vary non-linearly with TC activity, and the importance of some predictors changes with location. To account for spatial variability, past studies use heuristics or thresholds to cap the magnitudes of such variables and reduce spatial sensitivities. For example, larger absolute vorticities exist at higher latitudes, which would suggest a greater likelihood of TC genesis given its importance for genesis efficiency. Past works cap the absolute vorticity so that once a certain amount is reached, there is no increase in genesis probability. However, this practice introduces limitations in a non-stationary climate and may be subjective. As a result, we aim to improve the objectivity and spatial awareness of TCGIs and resolve non-linear relationships with TC activity.

Here, we use the ECMWF fifth-generation atmospheric reanalysis product (ERA5) and the International Best Track Archive for Climate Stewardship (IBTrACS) in the North Atlantic Basin from 1950–2004 to train a random forest model that serves as a TCGI and data from 2005–2022 for testing. We find that including latitude and longitude as predictors can improve the random forest model’s performance when evaluated using Pearson’s correlation. There are indications that the latitude input can help mitigate false positives at high latitudes resulting from the absolute vorticity issue. Below are four examples of predicted TC genesis values (shaded) from the testing dataset and observed TC genesis locations (red “x”) for August 2010, 2012, 2013, and 2019. In these examples, we observe model skill using a random forest in identifying concentrated regions of TC genesis. Additional variables being considered in the spatially aware TCGI include saturation deficit and/or potential intensity. Note that we also evaluate the model by replacing saturation deficit with column relative humidity, however, both variables are not used concurrently due to covariance. The creation of this spatially-aware index will help independently distinguish critical environmental variables for TC genesis in current and future climates while reducing the impact of climatological spatial variations.