Numeric Image Features for Detection of Aurora.

The electromagnetic coupling of the solar wind, Earth's magnetic field, and the upper atmosphere allows us to study the near-Earth space phenomena by monitoring the auroral displays in the polar regions. Ground-based networks facilitate spatial and temporal resolutions that are not possible with satellite instruments-they also produce enormous amounts of data to be stored and processed. While automated image analysis methods for auroral research are beginning to emerge, the normal approach is to visually examine images and then manually label and sort the data. We revisit a key question concerning the existence of aurora in an image: Not all images contain auroral light, and the visibility to the upper atmosphere depends on cloud cover. Detection of aurora is a fundamental step to limit further processing to only those images that are of interest. We quantitatively evaluated a selection of numeric image features that have been used in earlier studies and assess a brightness-invariant feature. We achieved error rates around 6%-8% with subsecond execution times. To the best of our knowledge, we are the first to report results in classifying auroral images where the Moon is allowed to be visible.