Speaker
Description
Accurate real-time prediction of coronal mass ejection (CME) arrivals is essential for mitigating the effects of space weather events on Earth. With the increasing volume of heliospheric imager (HI) data, manual detection and tracking of CMEs is becoming increasingly time-consuming. To address this challenge, we developed the Solar Transient Recognition Using Deep Learning (STRUDL) model, a convolutional neural network (CNN)-based pipeline for automatic CME detection and tracking in STEREO-A HI data. While STRUDL has shown promising results on science-quality HI data, its suitability for real-time applications has not yet been demonstrated.
In this study, we apply STRUDL to both real-time beacon data and enhanced beacon (E-beacon) data—a higher-resolution, higher-cadence version of beacon data —to assess model performance in an operational context. Using data from June 2024 to May 2025, we automatically segment CMEs and extract their time-elongation profiles. A threshold-based alert system is used to flag potential events of interest, producing an initial list of CMEs. Each detected event is then manually tracked to provide a high-quality reference dataset.
We use the STRUDL tracks obtained from the beacon and E-beacon data products, as well as the manual tracks, as input for ELEvoHI, a drag-based CME propagation model suitable for usage with HI data. All runs use consistent input parameters from the DONKI catalog to isolate the impact of the tracking method. We assess arrival time error, prediction lead time, and the overall detection performance in terms of true/false positives and false negatives over the study period. This work tests the feasibility of an end-to-end, real-time CME forecasting pipeline using HI data—critical for future missions such as ESA’s Vigil, which will rely on HI observations for continuous, operational space weather monitoring from the Sun–Earth L5 point.
| Do you plan to attend in-person or online? | In-person |
|---|
