Retrieving Tropical Cyclone Intensity from Climate Reanalysis using Deep Learning

Abstract

Traditional methods for improving tropical cyclone (TC) intensity from climate model outputs or projections have primarily relied on either dynamical or statistical downscaling. With recent advances in deep learning (DL) techniques, an important question is how DL can provide an alternative approach for enhancing TC intensity and structure retrieval from climate data. Using a common DL architecture based on convolutional neural networks (CNN) and a set of key environmental features relevant to TCs, we show that TC intensity and size can be effectively retrieved from climate reanalysis data without requiring super-resolution enhancement as in previous studies, even when applied to coarse-resolution climate data. This approach allows for retrieving TC intensity metrics and size that are dynamically constrained by the data, rather than estimating these quantities independently. Our results highlight that TC intensity and size are governed not only by TC internal processes but also by local environments during TC development for which DL models can learn and capture. The performance of our DL model depends on several factors such as season, the stage of TC development, or ocean basins, with root-mean-square errors ranging from $\approx$8-10 m s${}^{-1}$ for the maximum 10-m wind, $\approx$10-13 hPa for minimum central pressure, and $\approx$13-21 km for the radius of the maximum wind. Although these errors are better than any direct vortex detection or statistical downscaling methods applied to the same data, their wide ranges also suggest that a 0.5${}^{\circ }$-resolution climate data may contain limited TC information for DL models to learn from, regardless of model optimizations or architectures. Possible improvements and challenges in addressing the lack of fine-scale TC information in coarse-resolution climate reanalysis datasets are discussed.