Climate Downscaling of Tropical Cyclone Intensity using Deep Learning

TL;DR

The paper demonstrates that a CNN-based TCNN can infer tropical cyclone intensity and inner-core structure from coarse-grained climate data, outperforming direct vortex-tracking on the grid. By training on MERRA-2 fields and IBTrACS labels, the model captures VMAX, PMIN, and RMW with RMSEs and MAEs competitive with or better than traditional downscaling approaches, even at $0.5^{\circ}$ resolution. Key findings show that domain size, kernel size, data sampling, and moisture-related channels strongly influence performance, with environmental signals enabling downscaling beyond visible inner-core details. While promising, the study also highlights limitations due to coarse resolution and imperfect ground-truths, suggesting transfer learning to higher-resolution reanalyses and ongoing exploration of DL architectures to further improve TC downscaling.

Abstract

Traditional methods for enhancing 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, a natural question is whether DL can provide an alternative approach for improving TC intensity estimation from climate data. Using a common DL architecture based on convolutional neural networks (CNN) and selecting a set of key environmental features, we show that both TC intensity and structure can be effectively downscaled from climate reanalysis data as compared to common vortex detection methods, even when applied to coarse-resolution (0.5-degree) data. Our results thus highlight that TC intensity and structure are governed not only by its internal dynamics but also by local environments during TC development, for which DL models can learn and capture beyond the potential intensity framework. The performance of our DL model depends on several factors such as data sampling strategy, season, or the stage of TC development, with root-mean-square errors ranging from 3-9 ms$^{-1}$ for maximum 10 m wind and 10-20 hPa for minimum central pressure. Although these errors are better than direct vortex detection methods, their wide ranges also suggest that 0.5-degree 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 will be discussed.

Figures (11)

• Figure 1: The TCNN architecture for downscaling TC intensity from gridded climate data, with hyperparameter information noted for each layer and corresponding operation.
• Figure 2: Comparison of the predicted VMAX (unit, kt) as obtained from the TCNN model with the single-output design (blue), multiple-output design (red), directly vortex tracking on the model grid (green), and the observed intensity from the best track (black) for the randomly-sampled test set in the form of a) box plots, and b) a scatter plot. The thin red line in (b) denotes the perfect forecast.
• Figure 3: Similar to Fig. \ref{['fig:wind_speeds']} but for the minimum central pressure PMIN.
• Figure 4: Pressure wind relationship as obtained from the best track data (red) and that obtained from the TCNN model (blue) with (a) a multiple-output design, and (b) a single-output design. Solid thin lines denote the best quadratic fitting.
• Figure 5: Similar to Fig. \ref{['fig:wind_speeds']} but for the radius of maximum wind RMW.