Automatic Burst Detection in Solar Radio Spectrograms using Deep Learning: deARCE method

Abstract

We present in detail an automatic radio burst detection system, based on the AlexNet convolutional neural network, for use with any kind of solar spectrograms. A full methodology for model training, performance evaluation and feedback to the model generator has been developed with special emphasis on: i) robustness tests against stochastic and overfit effects; ii) specific metrics adapted to the imbalanced nature of the solar burst scenario; iii) tunable parameters for probability threshold optimization; iv) burst coincidence cross match among e-Callisto stations and with external observatories (NOAA-SWPC). The resulting neural network configuration has been designed to accept data from observatories other than e-Callisto --either ground- or satellite-based. Typical False Negative and False Positive Scores in single-observatory mode are, respectively, in the 10--16\% and 6--8\% ranges, which improve further in cross-match mode. This mode includes new services (deARCE, Xmatch) allowing the end scientist to check at a glance if a solar radio burst has taken place with a high level of confidence.