ByteStorm: a multi-step data-driven approach for Tropical Cyclones detection and tracking

TL;DR

ByteStorm addresses tropical cyclone tracking through a threshold-free, data-driven framework that combines two DL models for center detection (classification and localization) with the BYTE multi-object tracking algorithm to assemble coherent TC tracks. It relies on just RV850 and MSLP from ERA5 and IBTrACS labels, trained on a patch-based representation over the ENP and WNP basins. The approach delivers high detection performance, strong inter-annual variability correlations, and smoother tracks than traditional deterministic trackers, while remaining computationally efficient. These results demonstrate the potential of integrating deep learning with computer vision for fast, scalable TC diagnostics in climate science and operational contexts.

Abstract

Accurate tropical cyclones (TCs) tracking represents a critical challenge in the context of weather and climate science. Traditional tracking schemes mainly rely on subjective thresholds, which may introduce biases in their skills on the geographical region of application. We present ByteStorm, an efficient data-driven framework for reconstructing TC tracks without threshold tuning. It leverages deep learning networks to detect TC centers (via classification and localization), using only relative vorticity (850 mb) and mean sea-level pressure. Then, detected centers are linked into TC tracks through the BYTE algorithm. ByteStorm is evaluated against state-of-the-art deterministic trackers in the East- and West-North Pacific basins (ENP and WNP). The proposed framework achieves superior performance in terms of Probability of Detection ($85.05\%$ ENP, $79.48\%$ WNP), False Alarm Rate ($23.26\%$ ENP, $16.14\%$ WNP), and high Inter-Annual Variability correlations ($0.75$ ENP and $0.69$ WNP). These results highlight the potential of integrating deep learning and computer vision for fast and accurate TC tracking, offering a robust alternative to traditional approaches.

Figures (10)

• Figure 1: Overview of the ByteStorm approach. The input maps, consisting of RV850 and MSLP drivers, are divided into non-overlapping patches and processed by two models to (i) detect the TCs and (ii) locate their position within the patch. The BYTE method is then applied to link TC center detections into coherent spatio-temporal tracks.
• Figure 2: Architecture of the VGG-like models used in this work. A series of 6 CNN blocks is used as convolutional backbone. The CNN model is followed by a series of 4 Linear blocks that decode the spatial information coming from the CNN into a representation of TC presence/absence in the case of Classification model or the location of the TC eye in the case of Localization model. $f(x)$ is the ReLU activation function.
• Figure 3: Probability of Detection (POD) - higher is better - and False Alarm Rate (FAR) - lower is better - of the deterministic TC trackers compared with ByteStorm, on the ENP (top) and WNP (bottom) basins.
• Figure 4: De-Trended Inter-Annual Variability (IAV) of deterministic TC trackers compared to ByteStorm on ENP (top) and WNP (bottom) basins. ByteStorm is shown with a red thicker line.
• Figure 5: Track Duration is represented as the number of TCs (y-axis) lasting for a certain amount of days (x-axis). All the tracking schemes closely follow IBTrACS historical observations.