Objective Detection of Tropical Cyclone Formation from Remotely-sensed Ocean Surface Wind Using Machine Learning Approaches

Abstract:

While recent microwave remote sensing has provided opportunities to retrieve ocean surface wind in many developing and non-developing disturbances, current application to detection of tropical cyclone (TC) at genesis stage remains qualitative and subjective. For the best use of the microwave-sensed low-level dynamics, this study develops new objective detection techniques of TC genesis with machine learning approaches from WindSat data. Dynamical and hydrological factors related to TC genesis have been defined from 325 developing and 974 non-developing Windsat wind/precipitation images for 2005-2009. In particular, degrees of symmetry in the ocean-surface circulation are quantified from circular variances (CVs), to measure a spread of wind angles, and the degree of organization of strong wind (rainfall) is newly defined using a Spatial Pattern Analysis Program for Categrical Maps called as FRGSTATS. Also, system strength and convective vigor are represented by the averages of wind speeds and rainfall, respectively.

Applying the invented WindSat-derived predictors (2005-2007) to two machine learning approaches, decision tree (DT) and random forecast (RF), constructs objective detection techniques of TC genesis. Both results commonly show the symmetry and the intensity of ocean surface circulation are the most important than other predictors. Despite of inherent sampling issues of the polar orbiting satellite, hindcast validation with DT (RF) for 2008-2009 years show that a positive rate of detection is approximately 95.3% (83.7%) with false alarm rate 28.4% (11.1%). Accordingly, this study emphasizes new capability of the microwave-sensed dynamical information for objective detection of TC genesis.