Multi-Resolution Analysis of the Convective Structure of Tropical Cyclones for Short-Term Intensity Guidance

TL;DR

This work tackles 24-hour tropical cyclone intensity guidance by leveraging a multi-resolution analysis (MRA) of GOES-IR satellite imagery. It uses a discrete wavelet transform to produce sparse, multi-scale convective-structure features, enabling ML models to link structural patterns to rapid intensification while maintaining interpretability. The approach yields improved RI nowcasting performance and data compression, with interpretability aided by class activation maps in the wavelet space, and outlines a pathway to generative, wavelet-space forecasts using transformer-based autoregression. By compressing the data by roughly a factor of $1/20$ and grounding predictions in physically meaningful structures, the method offers practical potential to enhance lead times and forecaster trust for short-term TC intensity guidance.

Abstract

Accurate tropical cyclone (TC) short-term intensity forecasting with a 24-hour lead time is essential for disaster mitigation in the Atlantic TC basin. Since most TCs evolve far from land-based observing networks, satellite imagery is critical to monitoring these storms; however, these complex and high-resolution spatial structures can be challenging to qualitatively interpret in real time by forecasters. Here we propose a concise, interpretable, and descriptive approach to quantify fine TC structures with a multi-resolution analysis (MRA) by the discrete wavelet transform, enabling data analysts to identify physically meaningful structural features that strongly correlate with rapid intensity change. Furthermore, deep-learning techniques can build on this MRA for short-term intensity guidance.

Figures (2)

• Figure 1: Left: CNNs trained with our wavelet coefficients (solid blue) lead to much better TC intensity predictions, as indicated by a larger area under the ROC curve, than when working directly with the original GOES imagery (solid red). Right: Adding SHIPS variables marginally improves the wavelet results (dashed blue), but is necessary for GOES (dashed red after adding SHIPS) to be competitive with wavelet results without SHIPS (solid blue).
• Figure 2: Panel A: Example images at time $t=0$ (i.e., the last frame in a sequence $\mathbf{S_{\leq 0}}$) that is inputted to the respective CNNs to nowcast TC intensities for Hurricane Eta (2020) which underwent RI at $t=0$. Left: The high-resolution GOES-IR imagery. Right: The thresholded wavelet coefficients used as model inputs. In this case, $j=3$ represents the coarsest scale, and $j=1$ represents the finest scale used. Panel B:Example of class activation maps (CAMs) for the inputs at $t=0$ seen in panel A. Left: The CNN trained on the original GOES-IR imagery had a hard time discerning structural aspects of the storm, instead applying weight at the input's edges. Right: For the wavelet approach, the CNN was able to identify unique and specific structural aspects of the storm correlated with intensification, such as the storm's curvature and cloud coverage. Specific convective features were identified at all resolutions.