Decreasing False-alarm Rates in CNN-based Solar Flare Prediction Using SDO/HMI Data

A hybrid two-stage machine-learning architecture that addresses the problem of excessive false positives (false alarms) in solar flare prediction systems is investigated. The first stage is a convolutional neural network (CNN) model based on the VGG-16 architecture that extracts features from a temporal stack of consecutive Solar Dynamics Observatory Helioseismic and Magnetic Imager magnetogram images to produce a flaring probability. The probability of flaring is added to a feature vector derived from the magnetograms to train an extremely randomized trees (ERT) model in the second stage to produce a binary deterministic prediction (flare/no-flare) in a 12 hr forecast window. To tune the hyperparameters of the architecture, a new evaluation metric is introduced: the "scaled True Skill Statistic."It specifically addresses the large discrepancy between the true positive rate and the false positive rate in the highly unbalanced solar flare event training data sets. Through hyperparameter tuning to maximize this new metric, our two-stage architecture drastically reduces false positives by ≈48% without significantly affecting the true positives (reduction by ≈12%), when compared with predictions from the first-stage CNN alone. This, in turn, improves various traditional binary classification metrics sensitive to false positives, such as the precision, F1, and the Heidke Skill Score. The end result is a more robust 12 hr flare prediction system that could be combined with current operational flare-forecasting methods. Additionally, using the ERT-based feature-ranking mechanism, we show that the CNN output probability is highly ranked in terms of flare prediction relevance.