(News & views on the paper by Ru-Shi Lan et al., RAA, 2012, vol.12, 1191-1196)
Automated flare prediction using the AdaBoost algorithm
(Department of Physics and CSPAR, University of Alabama in Huntsville, AL, USA)
Our Sun is a powerful particle accelerator. Charged particles (including electrons, protons and various ions) are accelerated in solar flares and coronal mass ejections (CMEs). Propagating along interplanetary magnetic field lines from the Sun to the Earth, these high energy particles can pose an extreme deadly radiation dose to astronauts. In addition, they can affect electronics onboard spacecraft as well as commercial flights over the polar regions. They are one major concern of Space Weather studies. Being able to predict the occurrence of solar flares, therefore, has attracted a lot of research attention.
Many mechanisms have been proposed to explain solar flares. These include, for example, flux emergence and cancellation (e.g. Gan et al. 1993; Zhang et al. 2001), kink instability of coronal flux ropes (Sakurai 1976; Li & Gan 2011), and magnetic reconnection (Forbes et al. 2006; Fang et al. 2010 and 2012). In essence, all of these mechanisms examine how the magnetic field configuration varies such that a sudden release of the free magnetic energy is possible. As such, it is possible that certain features of the pre-eruption magnetic field can be used to predict/forecast flares.
It is not an easy task, considering that it encompasses many fields including solar physics, probability theory, computer graphics, etc. Flare predictions include two major aspects: one is the extraction of predictors and the other is the prediction algorithm. The work of Lan et al. in this letter discussed a particular method which uses the AdaBoost Algorithm (Freund et al. 1997) to predict the occurrence of solar flares. The procedure uses three predictors which are extracted from the photospheric magnetograms and by evaluating these predictors together, the probability of the occurrence of a flare larger than a certain level over a 24-hour period after the magnetogram was recorded is obtained. Comparing to another recently developed method [Yuan et al. 2010], which combines an ordinal Logistic Regression (LR) model and a Support Vector Machine (SVM) classifier, the prediction method presented in this letter is more accurate at predicting large flares but less accurate at predicting small flares.