A Systematic Framework for Data Augmentation for Tropical Cyclone Intensity Estimation Using Deep Learning

Abstract Given the exponential relationship between the intensity and the expected damage of tropical cyclones, accurately estimating their intensity from satellite images is a crucially important area of research. Yet, the imbalance and limited size of available data sets significantly hinder model training and generalization capabilities, especially for state‐of‐the‐art deep learning models. Therefore, it is standard in this field to use data augmentation, a family of transformations to increase the size and variety of a given data set. However, the principles behind the usage of these techniques for the estimation of the intensity of tropical cyclones has been largely unexamined. In this paper, we introduce a framework for establishing how much augmentation to apply and which techniques to use. To determine the ideal amount of augmentation, we use a modified Gini coefficient to understand how the augmentation will affect the imbalance in the data set, and we find that there is an upper bound on how much augmentation should be done (for our data set, 50 times per image). Our results also indicate that data augmentation works best when used to reduce the amount of imbalance in a data set rather than uniformly over the entire data set, as is typically done in the tropical cyclone intensity estimation literature. We then devise a backward elimination feature selection algorithm to determine which augmentation techniques work best. Our findings suggest that all augmentation techniques are effective for the estimation of tropical cyclones intensity, including random erasing, which we were the first to implement in this context.