Improving and Interpreting Flare Prediction with Spatial Statistics Analysis of the Magnet Field Data

• Activity Number: 81 - Statistical and Machine Learning Efforts on Solar Flare Predictions I
• Type: Topic-Contributed
• Date/Time: Monday, August 9, 2021 : 10:00 AM to 11:50 AM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Hu Sun* and Ward Manchester and Yang Chen
• Companies: University of Michigan, Ann Arbor and University of Michigan, Ann Arbor and University of Michigan
• Keywords: flare prediction; interpretability; spatial statistics; polarity inversion line; HMI data

Abstract

Flare prediction and classification tasks have seen great progress recently with the growing availability of data and the application of advanced machine learning models. But the great performances of many models rarely allows direct interpretability. Even with interpretable input features, it is hard to relate the spatial-temporal distributions of active regions features to the prediction scores. In our study, we undertake a strong-weak flare classification task with new features constructed from spatial statistics techniques, which describes the spatial distributions of vertical magnet component captured by the Helioseismic Magnetic Imager (HMI). These new spatial features are shown to provide an improvement of flare classification performance beyond that obtained using Spaceweather HMI Active Region Patch (SHARP) parameters alone, indicating the extra predictive signals hiding in the pixel-level spatial distributions. Visual illustrations of the spatial statistics features based on several case studies further reveal that these newly derived features have direct physical interpretations.