Abstract
We present a metric for detecting clouds in auroral all-sky images based
on single-wavelength keograms made with a collocated meridian
spectrograph. The coefficient of variation, the ratio of the sample
standard deviation to the sample mean taken over viewing angle, is the
metric for cloud detection. After calibrating and flat-field correcting
keogram data, then excluding dark sky intervals, the effectiveness of
the coefficient of variation as a detector is tested compared to true
conditions as determined by Advanced Very High Resolution Radiometer
(AVHRR) satellite imagery of cloud cover. The cloud mask, an index of
cloud cover, is selected at the corresponding nearest time and location
to the site of a meridian spectrograph at Poker Flat Research Range
(PFRR). We use events that are completely cloud-free or completely
cloudy according to AVHRR to compute the false alarm and missed
detection statistics for the coefficient of variation of the greenline
557.7 nm emission and of the redline 630.0 nm emission. For training
data of the years 2014 and 2016, we find a greenline threshold of 0.51
maximizes the percent of events correctly identified at 75%. When
applied to testing data of the years 2015 and 2017, the 0.51 threshold
yields an accuracy of 77%. There is a relatively shallow and wide
minimum of mislabeled events for thresholds spanning about 0.2 to 0.8.
For the same events, the minimum is narrower for the redline, spanning
roughly 0.3-0.5, with a threshold of 0.46 maximizing detector accuracy
at 78-79%.