Integrated and Interpretable Solar Flare Forecast with AIA and HMI Imaging Data: Application of the Spatial Transformer and Bayesian Neural Network
Abstract
Solar flare forecasting is a long-standing challenge in the astrophysics community, which has progressed significantly in recent years from the attention of the machine learning community supplied with high-resolution flare-related data from AIA and HMI instruments on the SDO spacecraft. Applications of modern deep learning techniques such as CNN and LSTM greatly boost the prediction accuracy but fall short on the interpretability, especially on how each AIA/HMI component contributes to the final prediction and where the flare precursors come from. In this project, we propose a shallow neural network, coupling spatial transformer and a bayesian neural network, to learn to predict strong solar flares with a combination of both AIA and HMI data. The new framework first learns to transform each AIA and HMI imaging component to a "canonical" orientation, and then learns a global filter to compress the transformed images to a probability score. The new framework allows direct interpretation of where the flare precursors come from in the AIA and HMI images, and also allows inferences on how different AIA and HMI channels interact in providing the prediction results. Our model strikes the balance between decent prediction accuracy and high interpretability while also providing a framework that accommodates data of different modalities, such as AIA, HMI and even the SHARP quantities, in the same machine learning model.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMNG42A..03S