B Longarela, R Roman, F. Buitrago, D Gonzalez-Fernandez, J Gaton, C del Barrio, S Herrero-Anta, A de Diego-Fernandez, R Gonzalez, D Mateos, L Doppler, A Barreto, M Munoz, C Toledano, V Cachorro
Abstract
Cloud Cover (CC) is a key meteorological parameter for the Earth's radiative balance. Yet, continuous full-diurnal retrieval from ground-based all-sky cameras remains a challenge due to the limitations of classical thresholding methods, the lack of nighttime literature, and the absence of standardization in deep learning approaches. This work develops a methodology to retrieve CC fraction in oktas from all-sky images during day and nighttime conditions. To fulfill this goal, this study introduces the GOA-ALLSKY dataset as an open-source benchmark comprising both daytime and nighttime images classified into okta classes. Based on this dataset, a rigorous framework evaluates model robustness under training-like conditions, varying aerosol loads, and a combined geographical and minor instrumental shift. Although this spatial shift inherently degrades the performance of all evaluated models, a proposed few-shot domain adaptation strategy successfully recovers partially their predictive potential. A comparative analysis demonstrates that modern architectures (ConvNeXt v2, Swin, and EfficientNet) significantly outperform traditional backbones such as VGG. Among them, ConvNeXt v2 exhibits the strongest extrapolation capabilities, reaching an overall test exact accuracy of 69.31% and a
Keywords
Cloud cover / All-sky cameras / Neural networks / Geographical shift / Few-shot domain adaptation / Deep learning
Atmospheric Research
Volume 346, Number 109403
2027 February





