A Robust Deep Learning Framework for Prominence Detection through Composite Feature Representations

Technical Summary: A Robust Deep Learning Framework for Prominence Detection through Composite Feature Representations

Problem Statement
Solar prominences are dynamic, cool plasma structures suspended in the hot solar corona. Their detection is critical for space weather forecasting, particularly regarding eruptions linked to coronal mass ejections. While deep learning offers a pathway to process the vast, high-cadence datasets provided by missions like the Solar Dynamics Observatory (SDO)/Atmospheric Imaging Assembly (AIA), previous attempts at automatic prominence detection have struggled. Specifically, models trained on standard SDO/AIA 304 Å channel data have exhibited poor performance, often failing to distinguish prominences from hot coronal emission (specifically the Si XI 303.3 Å line) or suffering from contamination by active regions. Furthermore, existing studies have reported low accuracy rates (around 50%) for prominence detection compared to other solar features, largely due to the low signal-to-background contrast and structural ambiguity of prominences in the 304 Å pass-band.

Methodology
The authors developed and optimized a deep learning framework based on the You Only Look Once (YOLOv5) object detection architecture to address these challenges. The study utilized a labeled dataset of full-disk SDO/AIA 304 Å images (2010–2017) originally created by Baek et al. (2021).

The research proceeded in two main phases:

1. Baseline Model (304 Colormap): The authors first trained a YOLOv5 model directly on the standard JPEG-format 304 Å images. This served as a control to evaluate performance and identify potential biases.
2. Composite Feature Models: To mitigate the limitations of the standard colormap and instrument degradation, the authors constructed a new preprocessing pipeline. This pipeline involved:
   • Calibrating raw FITS data to level 1.5 to correct for instrument degradation.
   • Applying the Wavelet-Optimized Whitening (WOW) technique to enhance fine structural details and contrast while suppressing noise.
   • Creating a three-channel input tensor for the neural network. The authors tested two variants:
     • Composite Model: Channels consisted of (i) full-disk grayscale, (ii) full-disk enhanced corona (WOW-processed), and (iii) disk-removed grayscale.
     • Composite Disk-removed Model: All three channels contained disk-removed data representations (with the first channel log-scaled for distinction).

The models were trained using Stochastic Gradient Descent (SGD) with momentum, employing on-the-fly data augmentations (rotation, flipping, scaling) to improve robustness. Performance was evaluated using standard object detection metrics: Precision, Recall, F1-score, and mean Average Precision (mAP@50 and mAP@50-95).

Key Contributions

• Identification of Colormap Bias: The study demonstrated that models trained on standard 304 Å colormaps exhibit a strong bias toward the red channel of the RGB representation, relying on color artifacts rather than physically meaningful shape or texture features. This bias was shown to degrade performance on newer, degraded images.
• Composite Feature Representation: The authors introduced a novel multi-channel input strategy combining grayscale, WOW-enhanced, and disk-removed representations. This approach physically enhances prominence structures and mitigates the obscuring effects of hot coronal emission and active regions.
• Cross-Instrument Generalization: The framework was tested on data from the Solar Ultraviolet Imager (SUVI) onboard the GOES-R series, demonstrating the model's ability to generalize across different instruments despite differences in field of view and image rotation.
• Bias Analysis in Ground Truth: The paper highlights that "false positives" identified by the model often correspond to valid, unlabeled prominences that were missed in the original ground-truth labeling due to the limitations of the standard 304 Å view. This finding suggests that standard metrics may underestimate model performance when the ground truth itself is incomplete.

Results

• Performance Metrics: The Composite Model achieved the best overall performance with a Recall of 78% and an mAP@50 of 0.749. This represents a significant improvement over previous bounding box methods (e.g., SSD and Faster R-CNN), which reported accuracies around 50–59% on the same dataset.
• Effective Precision: Visual analysis revealed that approximately 74% of the model's apparent false positives were actually valid, unlabeled prominences. When accounting for these "hidden" true positives, the effective precision of the Composite Model rose to 91%, and the Disk-removed variant to 93%.
• Model Comparison: While the Composite Disk-removed model showed slightly lower metric scores and slower convergence, it was more conservative. The standard 304 Colormap model, despite achieving a slightly higher mAP@50 (0.755) in training, suffered from channel dependency and failed to detect features obscured by coronal emission or active regions.
• Generalization: The Composite Model successfully detected prominences in SUVI data, identifying features that were visible in the enhanced processing but sometimes missed in standard AIA views, confirming its utility across instruments.

Significance and Claims
The paper claims to provide a reliable, physically motivated, and versatile deep learning framework for the automatic detection of solar prominences. The authors emphasize that their approach moves beyond the limitations of standard colormaps and single-channel inputs, offering a robust solution for space weather applications.

Crucially, the paper underscores the importance of dataset construction. It argues that biases in image selection (temporal imbalance due to instrument degradation) and labeling (incomplete detection of faint structures) directly propagate into model predictions and performance statistics. The authors propose that future robust datasets must:

1. Be selected uniformly across the solar cycle.
2. Avoid extraneous features like text artifacts.
3. Utilize multi-channel composite data or grayscale representations rather than arbitrary colormaps.
4. Adhere to consistent labeling standards, potentially re-labeling data using enhanced processing techniques (like WOW) to capture the full extent of prominence structures.

The study concludes that while the current model is a significant step forward, future work should focus on extending the framework to include active region classification and integrating physical property extraction methods to build comprehensive prominence catalogs.