[Paper Review] Flare forecasting and feature ranking using SDO/HMI data

Michele Piana, Cristina Campi+3 more|arXiv (Cornell University)|Dec 18, 2018

Solar and Space Plasma Dynamics4 citations

TL;DR

This paper presents a hybrid supervised/unsupervised machine learning method for solar flare forecasting using SDO/HMI vector magnetogram data, leveraging LASSO regression to predict flare occurrence and rank feature importance. The approach achieves strong predictive performance (AUC > 0.85 for C1+ flares) and identifies key magnetic properties—such as vertical current density and magnetic shear—that significantly influence flare likelihood.

ABSTRACT

We describe here the application of a machine learning method for flare forecasting using vectors of properties extracted from images provided by the Helioseismic and Magnetic Imager in the Solar Dynamics Observatory (SDO/HMI). We also discuss how the method can be used to quantitatively assess the impact of such properties on the prediction process.

Motivation & Objective

• To develop a machine learning framework for predicting solar flares using SDO/HMI vector magnetogram data.
• To quantify the contribution of individual magnetic features to flare prediction accuracy.
• To enable interpretable flare forecasting by ranking features based on their predictive impact.
• To integrate the method into the FLARECAST platform for operational space weather applications.
• To support the design of future solar observation instruments by identifying the most informative magnetic properties.

Proposed method

• Extracted 171-dimensional feature vectors from SDO/HMI full-disk vector magnetograms using SMART tools and FLARECAST pattern recognition.
• Applied LASSO regression to train a supervised model on historical data, solving the optimization problem: $\hat{\beta} = \arg\min_{\beta}(\|y - X\beta\|_2^2 + \lambda\|\beta\|_1)$.
• Used the LASSO solution $\hat{\beta}$ to compute theoretical labels $\hat{y} = X\hat{\beta}$ for subsequent clustering.
• Applied fuzzy clustering to the predicted labels to determine centroids for flare (label 1) and non-flare (label 0) classes.
• Classified new magnetograms by comparing $x_{\text{new}}^T \hat{\beta}$ to the cluster centroids for binary prediction.
• Ranked features by the magnitude of their LASSO coefficients $\hat{\beta}_i$, quantifying each feature's impact on prediction.

Experimental results

Research questions

• RQ1 Which magnetic properties extracted from SDO/HMI vector magnetograms are most predictive of subsequent solar flares?
• RQ2 How does the hybrid LASSO-based method compare in predictive performance to other established machine learning models for flare forecasting?
• RQ3 Can the LASSO regularization technique effectively identify and rank the most relevant features for flare prediction while ensuring model interpretability?
• RQ4 To what extent can the feature ranking results inform the design of future solar observation instruments?
• RQ5 How stable are the feature rankings across multiple random train-test splits of the data?

Key findings

• The hybrid LASSO method achieved a True Skill Statistic (TSS) of 0.68 ± 0.05 and a Heidke Skill Score (HSS) of 0.52 ± 0.06 for predicting C1+ flares within 24 hours.
• The method outperformed other models (SVC, random forest, logit-based hybrid) in terms of TSS and HSS, with a Probability of Detection (POD) of 0.72 ± 0.04.
• The top-ranked features included vertical current density, magnetic shear, and horizontal current density, indicating their strong predictive relevance.
• Feature ranking was stable across 100 random training set realizations, with the top 10 features maintaining consistent rank order.
• The average Area Under the ROC Curve (AUC) for C1+ flare prediction was 0.85 ± 0.03, indicating strong discriminatory power.
• The method was successfully integrated into the FLARECAST platform and is publicly accessible for open use and continuous updates.