Machine Learning Approach to Identifying Aerosol and Geophysical Variable Influence on Southwest Pacific Ocean Tropical Cyclone Development

This study evaluates the ability of machine learning algorithms to distinguish between tropical depression (TD) and tropical storm (TS) environments in the southwest Pacific Ocean (SWPO). TDs and TSs are identified based on a maximum sustained wind speed threshold ( 17 ms^-1). Various aerosol, thermodynamic, and dynamic variables are extracted nearest to each TD initiation point and each TS intensification point. The variables associated with each labeled sample are used to train a decision tree and random forest classifier. To achieve optimal performance, random forest hyperparameters are tuned with repeated 10-fold cross validation. Results using a testing data set show that the random forest approach more accurately distinguishes between TD and TS samples. An analysis of variable importance from a trained random forest model suggests that sea salt, 1000 mb relative humidity, sea surface temperature, 850 mb air temperature, 925 mb relative vorticity, and sulfate are among the most important distinguishing variables. The machine learning outcomes are verified using interpretations of the physical environment during active and inactive SWPO TC seasons. Each set of TD and TS cases was used to develop composites of important aerosol and geophysical features to facilitate geophysical comparisons during TD and TS environment. This work will advance the risk management strategies for northeastern Australia and other SWPO basin islands to control their TC related losses.