Automated Night Time Cloud Detection using Keograms

Global Navigation Satellite System (GNSS) L-band scintillations at high latitudes have been observed to be associated with auroral forms. A systematic survey of scintillations from an array of GNSS receivers at Poker Flat Research Range (PFRR), Alaska, has shown thousands of scintillations occurred from 2014-2019, with frequency of occurrence declining with the solar cycle. Before comparing these scintillation events with the presence of aurorae in all-sky images, it is necessary to identify the presence of clouds in the sky, which can obscure or scatter auroral light reducing location accuracy. useful to efficiently and automatically.

A keogram is summary representation of intensity at a given wavelength along a single swath (e.g., an elevation sweep at constant azimuth) in the sky as a function of time.
We propose a simple metric, based on keogram data, which are more compact than all-sky images, for the detection of cloud cover. The test statistic at a given keogram timestamp is the coefficient of variation, which is the ratio of the standard deviation of the intensity at all elevations to the mean intensity. When collocated with an all-sky imager, the keogram can then efficiently help determine useful time periods of all-sky image for analysis. Cloudy events may either be removed from consideration or set aside for later manual processing.

We test the effectiveness of the coefficient of variation as a cloud detection test statistic using a keogram sited at Poker Flat Research Range (PFRR). Using the 630.0 nm (red) and 557.7 nm (green) intensities, after calibration and flat-field correction, we compare the detection results with the NOAA Advanced Very High Resolution Radiometer (AVHRR) + High-resolution Infra-Red Sounder (HIRS) cloud properties data. For every day considered in 2014-2016, the cloud mask data from within 5 miles of PFRR are considered. The distributions of the test statistic differ for cloud-free periods and for cloudy conditions. This means that for a desired false alarm rate or missed detection rate, a threshold value for the coefficient of variation in the keogram images can be set. For thresholds of 0.25 for 557.7 nm and 0.4 for 630.0 nm, the percent of false alarms is around 2% and for missed detections of cloudy periods about 30%. Modeling the distributions of the test statistic, we can choose thresholds analytically to select a desired missed detection or false alarm rate. This approach enables us to systematically scan keograms collocated with all-sky images for cloud-free periods, which in future can then be cross-referenced with occurrences of L-band scintillation.