Advancing Solar Flare Prediction using Deep Learning with Active Region Patches

TL;DR

This work addresses solar flare prediction for $\geq M$-class events across the full solar disk by using limb-to-limb AR patches derived from LoS magnetograms and a sliding-window preprocessing that prioritizes regions with maximum total unsigned flux (USFLUX). It evaluates three lightweight architectures—ResNet34, MobileNet, and MobileViT—with augmentation and undersampling to handle extreme class imbalance, using a Composite Skill Score (CSS) defined as $CSS = \sqrt{TSS \cdot HSS}$ for ranking. MobileNet achieves CSS values around $0.51$ for AR patches within $\pm 30^{\circ}$ and $\pm 60^{\circ}$, $0.48$ for $\pm 90^{\circ}$, and $0.39$ for near-limb regions, indicating limb-to-limb prediction is feasible though with reduced performance toward the limb. The study reports superior CSS performance compared to several prior AR- and full-disk approaches within similar longitudinal ranges and provides an open-source pipeline (gsudmlab2024arpatchsfp) for reproducibility and further development.

Abstract

In this paper, we introduce a novel methodology for leveraging shape-based characteristics of magnetograms of active region (AR) patches and provide a novel capability for predicting solar flares covering the entirety of the solar disk (AR patches spanning from -90$^{\circ}$ to +90$^{\circ}$ of solar longitude). We create three deep learning models: (i) ResNet34, (ii) MobileNet, and (iii) MobileViT to predict $\geq$M-class flares and assess the efficacy of these models across various ranges of solar longitude. Given the inherent imbalance in our data, we employ augmentation techniques alongside undersampling during the model training phase, while maintaining imbalanced partitions in the testing data for realistic evaluation. We use a composite skill score (CSS) as our evaluation metric, computed as the geometric mean of the True Skill Score (TSS) and the Heidke Skill Score (HSS) to rank and compare models. The primary contributions of this work are as follows: (i) We introduce a novel capability in solar flare prediction that allows predicting flares for each ARs throughout the solar disk and evaluate and compare the performance, (ii) Our candidate model (MobileNet) achieves a CSS=0.51 (TSS=0.60 and HSS=0.44), CSS=0.51 (TSS=0.59 and HSS=0.44), and CSS=0.48 (TSS=0.56 and HSS=0.40) for AR patches within $\pm$30$^{\circ}$, $\pm$60$^{\circ}$, $\pm$90$^{\circ}$ of solar longitude respectively. Additionally, we demonstrate the ability to issue flare forecasts for ARs in near-limb regions (regions between $\pm$60$^{\circ}$ to $\pm$90 $^{\circ}$) with a CSS=0.39 (TSS=0.48 and HSS=0.32), expanding the scope of AR-based models for solar flare prediction. This advancement opens new avenues for more reliable prediction of solar flares, thereby contributing to improved forecasting capabilities.

Figures (7)

• Figure 1: An illustrative example of (a) Original raw input magnetogram of HMI AR patch corresponding to HARP number: 4781 (NOAA AR number: 12205) at timestamp 2014-11-06T18:00:00 UTC, (b) Bitmap corresponding to HMI AR patch in (a) showing the region of interest indicated by white pixels, (c) Final processed image of AR patch in (a) now sized to 512$\times$512, that is used to train our models.
• Figure 2: The process flow diagram of data processing pipeline used in this work. It shows a sequential pipeline for creating JPEG images from magnetogram rasters and corresponding bitmaps along with data augmentation pipeline given the label for the magnetogram patch. Boxes colored in green collectively defines our entire dataset.
• Figure 3: (a)Overall data distribution (b) Overall data partitioned into four tri-monthly partitions. Note: The height of the bars are in log scale.
• Figure 4: An illustrative example of (a) input magnetogram of HMI AR patch corresponding to HARP number: 4781 (NOAA AR number: 12205) at timestamp 2014-11-06T18:00:00 UTC, (b-f) five different augmentations applied to AR patch in (a).
• Figure 5: Overall data partitioned into train set (showing original, undersampled and augmented data counts) , validation set, and test set used in this study.