Beacon2Science: Enhancing STEREO/HI beacon data with machine learning for efficient CME tracking

TL;DR

The paper tackles the challenge of real-time CME forecasting with noisy, low-resolution STEREO/HI beacon data. It introduces Beacon2Science, a two-stage ML pipeline that denoises and upsamples beacon frames to create E-beacon and then interpolates to IE-beacon to achieve 40-minute cadence, bridging beacon and science data. Quantitative results show substantial gains in image fidelity (PSNR, SSIM) and CME-tracking accuracy, with MAE reductions from about 1.0 degree to roughly 0.50 degrees relative to science tracks, and near-real-time inference on GPUs. The approach enhances near-term space weather forecasting capabilities and is positioned to support upcoming missions like Vigil and PUNCH.

Abstract

Observing and forecasting coronal mass ejections (CME) in real-time is crucial due to the strong geomagnetic storms they can generate that can have a potentially damaging effect, for example, on satellites and electrical devices. With its near-real-time availability, STEREO/HI beacon data is the perfect candidate for early forecasting of CMEs. However, previous work concluded that CME arrival prediction based on beacon data could not achieve the same accuracy as with high-resolution science data due to data gaps and lower quality. We present our novel machine-learning pipeline entitled ``Beacon2Science'', bridging the gap between beacon and science data to improve CME tracking. Through this pipeline, we first enhance the quality (signal-to-noise ratio and spatial resolution) of beacon data. We then increase the time resolution of enhanced beacon images through learned interpolation to match science data's 40-minute resolution. We maximize information coherence between consecutive frames with adapted model architecture and loss functions through the different steps. The improved beacon images are comparable to science data, showing better CME visibility than the original beacon data. Furthermore, we compare CMEs tracked in beacon, enhanced beacon, and science images. The tracks extracted from enhanced beacon data are closer to those from science images, with a mean average error of $\sim 0.5 ^\circ$ of elongation compared to $1^\circ$ with original beacon data. The work presented in this paper paves the way for its application to forthcoming missions such as Vigil and PUNCH.

Figures (11)

• Figure 1: (Top): Comparison of the average pixel intensity of beacon compared to science running difference images. We indicate the $\beta$ choice of 2.5 for normalization. (Bottom): Comparison of two science and beacon pairs (a and b). Each science running difference is on the left, and the beacon is on the right.
• Figure 2: Summary of the training dataset, with the number of images obtained yearly from the HELCATS WP2 events
• Figure 3: The STEREO-A position in HEE coordinates for training set (dots), with the HI1 field-of-view of the selected test events (cones).
• Figure 4: Schematic overview of the first network from our pipeline. In the top panel, the generator network architecture is based on ResUNet, with PixelShuffle layers for super-resolution. In the bottom panel, the discriminator is a small convolutional neural network predicting images to be real or generated. We use respective feature space sizes of [64, 128, 256, 512] for the four residual blocks of the generator.
• Figure 5: Science J-maps at varying time-axis resolutions, using varying $\Delta_t$ for creating the running difference images. Higher time-axis resolution J-maps improve the visibility and separation of CME tracks even at higher $\Delta_t$