Automatic classification of colour all-sky images

A recently developed machine vision routine detects aurora in the colour image data with an accuracy higher than 90%. The method uses OpponentSIFT features coupled with Support Vector Machines (SVM) classification. The training has been performed with about 30000 manually labelled images from stations at Muonio and Kevo.

Sample animations

These animations include data from the station at Muonio on 13-14 November, 2012. The exposure time was 4 seconds an the image cadence 10 seconds. Wrongly classified images here mainly relate to very faint aurora or partly cloudy periods of observations.

For more information of the method and test results, see: Rao et al., "Automatic auroral detection in colour all-sky camera images", IEEE Selected Topics in Applied Earth Observations and Remote Sensing, DOI : 10.1109/JSTARS.2014.2321433, 2014.

Data requests: Noora Partamies