A novel machine learning framework for precursor identification and extreme hurricane prediction

Abstract

The frequency of extreme events is undoubtedly increasing, largely due to climate change. To mitigate these effects and anticipate their potential consequences, a reasonable strategy is to combine the fields of remote sensing (which provides data) and artificial intelligence (with its diverse models and algorithms). By leveraging a vast amount of properly processed data and employing complex machine learning models, we can gain insights into the evolution of events that may have catastrophic consequences. Our research efforts have led us to develop a series of artificial intelligence models specifically tailored to extreme cyclone cases, capable of identifying intrinsic non-linear relationships among certain variables associated with the cloud's system and the intensity categories that could serve as precursors to the development of extreme cyclones. Specifically, we have focused on predicting extreme hurricanes and Mediterranean cyclones (Medicanes). The primary goal of this Thesis is to anticipate the formation of these extreme events, providing timely information to decision-makers and enhancing our understanding of the potential factors that impact their development. Additionally, part of the analysis was dedicated to improving predictions for rapidly intensifying extreme events, which pose significant challenges for operational models. Therefore, the goal is twofold: 1. Enhance knowledge about the evolving precursors linked to extreme cyclone cloud systems, enabling better anticipation of this specific phenomenon; 2. Develop innovative AI-based tools capable of revealing the intricate relationships between an optimal combination of precursors and the maximum development of extreme cyclones. Furthermore, the aim was to design models that exhibit flexibility to accommodate the inclusion of new variables or intensity categories for future exploration. These models serve as foundational building blocks from an AI perspective. This work resulted in three peer-reviewed publications that successfully addressed our goals, and includes additional contributions derived from related projects and initiatives.