Segmentation and Tracking of Eruptive Solar Phenomena with Convolutional Neural Networks

TL;DR

The paper tackles the challenge of automatically segmenting and tracking solar eruptive phenomena in high-dynamic-range EUV imagery. It combines the Wavetrack wavelet-based feature masks with CNN-based segmentation (U‑Net–like architectures) and a hyperparameter search to optimize performance on SDO/AIA data. Key contributions include data-driven training using synthetic augmentation and curriculum-style strategies, demonstration on a major CME event, and integration of velocity-field estimation to validate and refine feature tracking. The approach offers scalable, automation-friendly segmentation suitable for operational heliophysics pipelines and potential onboard processing, advancing our ability to analyze solar eruptions and their space-weather impacts.

Abstract

Solar eruptive events are complex phenomena, which most often include coronal mass ejections (CME), CME-driven compressive and shock waves, flares, and filament eruptions. CMEs are large eruptions of magnetized plasma from the Sun's outer atmosphere or corona, that propagate outward into the interplanetary space. Over the last several decades a large amount of remote solar eruption observational data has become available from ground-based and space-borne instruments. This has recently required the development of software approaches for automated characterisation of eruptive features. Most solar feature detection and tracking algorithms currently in use have restricted applicability and complicated processing chains, while complexity in engineering machine learning (ML) training sets limit the use of data-driven approaches for tracking or solar eruptive related phenomena. Recently, we introduced Wavetrack - a general algorithmic method for smart characterization and tracking of solar eruptive features. The method, based on a-trous wavelet decomposition, intensity rankings and a set of filtering techniques, allows to simplify and automate image processing and feature tracking. Previously, we applied the method successfully to several types of remote solar observations. Here we present the natural evolution of this approach. We discuss various aspects of applying Machine Learning (ML) techniques towards segmentation of high-dynamic range heliophysics observations. We trained Convolutional Neural Network (CNN) image segmentation models using feature masks obtained from the Wavetrack code. We present results from pre-trained models for segmentation of solar eruptive features and demonstrate their performance on a set of CME events based on SDO/AIA instrument data.

Figures (9)

• Figure 1: U-NET architecture in general. Each gray box corresponds to a multi-channel feature map. (The number of channels is not specified, as it would depend on the network setup in each specific case)
• Figure 2: Five separate example CBF events from the initial dataset used for engineering of training set. May 11, 2011; June 07, 2011; December 12, 2013; December 26, 2013, May 07, 2021. Wavetrack-obtained masks were projected over inverted-colormap grayscale AIA 193Å images. Events and timesteps were selected to represent fronts of different shapes in various heliospheric latitude-longitude zones for the purposes of creating diverse training set with further engineering of a synthetic training data through data augmentation routine.
• Figure 3: Rotation argumentation example for the two events from the training set. Rotated BD images are shown on the first and third row, with corresponding feature masks (ground truth) in the second and forth rows, respectively. Columns correspond to random angles chosen by the algorithm during one of the training steps on synthetic data.
• Figure 4: Compact CNN model structure. It is based on a hyper-parameter search based on segmentation performance evaluation on BD images, [-50,150] absolute intensity threshold interval applied. Due to its vertical size the plot was split vertically into 2 columns for the purpose of more compact visual representation. Thin black arrow lines denote skip connections. The legend below the architecture describes the structure and dimensionality principle of the encoder and decoder blocks used.
• Figure 5: Large CNN model structure. It is based on a hyper-parameter search based on segmentation performance evaluation on BD images with no threshold applied. This more complex and deep network structure was generated as segmentation of a higher dynamic range images was attempted. Due to its vertical size the plot was split vertically into 2 fragments for the purpose of more compact visual representation. Thin black arrow lines denote skip connections. The legend below the architecture describes the structure and dimensionality principle of the encoder and decoder blocks used.