Synergy of Traditional Techniques and Convolutional Neural Networks for Classification of Cloudless Conditions by All-Sky Imagers

Abstract

This work explores the capabilities of our two methods for the determination of cloudless conditions from All-Sky Imager (ASI) data. For the first time, it was demonstrated that the combination of a well-established traditional computer vision technique based on the calculation of a Clearness index and a new data-driven method, that utilizes Convolutional Neural Networks, leverages the benefits of both methods and suppresses their individual disadvantages. The developed tool is reliable and efficient, allowing us to study the occurrence of clear sky conditions over an excellent astronomical location at El Leoncito Observatory in Argentina throughout the years 2006-2023. It was found that seasonal variations in cloudiness conditions are present over long-term measurements of 18 yr. The detected variations are connected to changes in seasons with a minimum of averaged clear sky hours in January. The minima of long-term variations occurred in the years 2016 and 2023 which implies the potential usage of ASI data in climatological studies. In total, 40,250 clear sky hours were identified. On average per year, ∼75% of all observed hours were without clouds. These periods can be used for further ionospheric and space physics studies with data obtained via particular ASI narrowband filters. The presented data-driven approach will be used also for these planned studies as it demonstrated a high potential for automation of image processing from ASI instruments.