Image feature extraction for solar flare prediction

Solar flare is the most violent solar activity which is the main driving source of space weather, so accurate prediction of flare occurrence in coming days would manner disaster treatment and protection. Due to detailed reasons of solar flares eruption are not clear in current field, so the prediction clues rely mainly on observing solar images. Many predictors have been used for solar flare prediction, mainly based on expert system or physical knowledge. In this paper, a system based on image information without prior physical knowledge for solar flare prediction is presented. The Magnetic field and texture distribution of active region, the largest sunspot group's fractal dimension, positive and negative areas and girth, extracted from SOHO/MDI longitudinal magnetograms are used in the model to describe the complexity of the photospheric magnetic field. Machine learning algorithms: C4.5 decision tree, CART tree and Bayesian network are employed to predict the flare level within 48 hours. It is concluded that the model trained by C4.5 decision tree could predict flare occurrence effectively.