Improved forecasting of extreme hurricane events by integrating spatio-temporal CNN-RF learning of tropical cyclone characteristics

Abstract

Assessing hurricane predictions in a changing climate is one of the most challenging weather forecast problems today. Furthermore, effectively integrating information-rich features that are specific to the growth of hurricanes proves to be a difficult task due to the anticipated nonlinear interactions during the spatio-temporal evolution of the tropical cyclone system. Consequently, the need arises for complex and nonlinear models to address this formidable scenario. In light of this, we introduce a novel framework that combines a Convolutional Neural Network with a Random Forest classification configuration. This innovative approach aims to identify the critical spatial and temporal characteristics associated with the formation of major hurricanes within the hurricane and surrounding regions of the Atlantic and Pacific oceans. Here, we demonstrate that the inclusion of these unprecedented spatio-temporal features extracted from brightness temperature data, along with the temperature and anatomical cloud properties of the system, results in an average improvement of 12% in the prediction of severe hurricanes, using the previous model version as a benchmark. This enhancement in the prediction accuracy extends up to 3 days in advance, considering both regions collectively. Although these innovative attributes may be relatively more costly to generate, it allows us to gain a more refined understanding of the intricate relationships between different spatial locations and temporal dynamics, leading to more efficient and effective solutions. This hybrid machine learning approach also offers adaptability, enabling the exploration of other suitable hurricane or environmental-related conditions, making it suitable for potential future applications.