A neural network to retrieve cloud cover from all‐sky cameras: A case of study over Antarctica

Abstract

We present a new model based on a convolutional neural network (CNN) to predict daytime cloud cover (CC) from sky images captured by all-sky cameras, which is called CNN-CC. A total of 49,016 daytime sky images, recorded at different Spanish locations (Valladolid, La Palma, and Izaña) from two different all-sky camera types, are manually classified into different CC (oktas) values by trained researchers. Subsequently, the images are randomly split into a training set and a test set to validate the model. The CC values predicted by the CNN-CC model are compared with the observations made by trained people on the test set, which serve as reference. The predicted CC values closely match the reference values within ±1 oktas in 99% of the cloud-free and overcast cases. Moreover, this percentage is above 93% for the rest of partially cloudy cases. The mean bias error (MBE) and standard deviation (SD) of the differences between the predicted and reference CC values are calculated, resulting in MBE = 0.007 oktas and SD = 0.674 oktas. The MBE and SD are also represented for different intervals of measured aerosol optical depth and Ångström exponent values, revealing that the performance of the CNN-CC model does not depend on aerosol load or size. Once the model is validated, the CC obtained from a set of images captured every 5 min, from January 2018 to March 2022, at the Antarctic station of Marambio (Argentina) is compared against direct field observations of CC (not from images) taken at this location, which is not used in the training process. As a result, the model slightly underestimates the observations with an MBE of −0.3 oktas. The retrieved data are analyzed in detail. The monthly and annual CC values are calculated. Overcast conditions are the most frequent, accounting for 46.5% of all observations throughout the year, rising to 64.5% in January. The annual mean CC value at this location is 5.5 oktas, with a standard deviation of approximately 3.1 oktas. A similar analysis is conducted, separating data by hours, but no significant diurnal cycles are observed except for some isolated months.