Deep-learning based simulation of tropical cyclone genesis in Northwest Pacific

Abstract

The frequency and spatial distribution of tropical cyclone genesis (TCG) plays a crucial role in assessing tropical cyclone (TC) activities and relevant hazards. However, the generation of TCG involves complex mechanisms that are correlated to the background environment, and there is still significant room for better describing the distribution patterns of TCG despite the great achievements that have been made via classic statistical techniques and dynamical/thermodynamical methods. This study utilizes deep learning (DL) technology to investigate TCG patterns, with the primary aim of developing more reasonable sampling models with better generalization performance and satisfactory accuracy. Two approaches are proposed. The first one uses Variational Auto-encoder (VAE) model for direct (or non-parametric) TCG simulation, while the second one employs Convolutional Neural Network (CNN) to further explore environmental effects. For the second approach, two specific strategies have been examined. The first strategy describes TCG as a function of large-scale environment parameters (such as sea surface temperature, vorticity, and vertical wind shear), and the other one establishes relationships between TCG and typical parameters of the environment at multiple altitudes. Multiple evaluation indexes are also proposed to quantify the performance of adopted techniques from the aspects of generalization and accuracy. Results demonstrate that the proposed DL models perform better than classic statistical methods across various functional aspects, particularly in terms of generalization performance. Meanwhile, the DL models have great potential in assessing the effects of climate change on TCG patterns, which is absent or weakened in classic simulation methods. In sum, the proposed TCG simulation methods can be used to facilitate the assessment of TC hazards effectively.