Identifying Developing Cloud Clusters Using Predictive Features

Abstract

Forecasters need better data-driven techniques using feature extraction to determine whether a cyclone will develop from a loosely organized cluster of clouds. Prior studies have attempted to predict the formation of tropical cyclones using numerical weather prediction models and satellite and radar data. However, refined observational data and forecasting techniques are not always available or accurate in areas such as the North Atlantic Ocean where data are sparse. In response, this research investigates the predictive features that contribute to a cloud cluster developing into a tropical cyclone without using dynamic models. Instead, it will only use global gridded satellite data which are readily available. Generally, an imbalance occurs in the classification process of cloud clusters since the number of non-developing cloud clusters is greater than the number of developing cloud clusters. Imbalanced data are an essential source of low performance in learning about rare events. To address this issue, the produced cloud cluster feature dataset is balanced by applying the Selective Clustering based Oversampling Technique (SCOT), which addresses data imbalance in a selective manner and can be used in many applications. In this research, the predictive features are identified based on the performance of separating developing and non-developing cloud clusters from the balanced feature dataset when using a standard classifier. The predictive features are identified only if the classification yields a geometric mean of at least 80 % and a Heidke Skill Score of at least 0.8.