GOA-ALLSKY: An All-Sky Image Dataset for Cloud Cover Classification in Oktas

Overview The GOA ALL-SKY dataset provides 7,467 annotated images for automated Cloud Cover (CC) classification in oktas (0-8), following WMO guidelines. The data was acquired through the GOA-SCAN network, managed by the Group of Atmospheric Optics of the University of Valladolid (GOA-UVa). It is specifically designed to address challenges in all-sky imaging, including nighttime monitoring, subjective observer bias, and model generalization across different climates and spectral responses. Instrumentation & Image Processing Sensor: All images were captured using OMEA-3C all-sky cameras (Alcor System) equipped with Sony IMX178 CMOS sensors and 180∘ fisheye lenses. Processing: Images were acquired as multi-exposure sequences (every 5 min by day; 2 min by night) and processed into High Dynamic Range (HDR) images to maximize detail in both bright and dark regions. Format: The final exported images are 8-bit RGB, cropped to a resolution of 2000×2000 pixels. Dataset Structure Subset Images Location Year Purpose Key Features Train 5400 Valladolid, Spain (41.66º N, 4.70º W, 710 m a.s.l.) 2024 Model training 300/class | 50% day/night. Validation 720 Valladolid, Spain (41.66º N, 4.70º W, 710 m a.s.l.) 2025 Train-like Distribution evaluation 80/class | 50% day/night. Reviewed by a multiple observers. Test 720 Tauche, Germany (52.21º N, 14.12º E, 120 m a.s.l.) 2024 Out of Distribution Evaluation 80/class | 50% day/night. Reviewed by a multiple observers. AOD Test 350 Granada, Spain (37.16◦ N, 3.61◦ W, 680 m a.s.l.) 2024-2026 To asses AOD impact on performance 50 images per AOD at 440 nm interval (0.0-0.1, 0.1-0.2, ... , 0.6>) High AOD Adaptation Subset97 97 Valladolid, Spain (41.66º N, 4.70º W, 710 m a.s.l.) 2024 Model few-shot domain adaptation All 97 images have an AOD at 440 nm > 0.3 Moonless Adaptation Subset 180 La Palma, Spain (28.49º N, 17.85º W, 630 m a.s.l.) 2021-2022 Model few-shot domain adaptation 20/class. Focus on moonless nighttime scenarios. Reviewed by a second observer. Data are organized as follows: <pre><code>GOA-ALLSKY/├── train/│ └── 0/ ... 8/ # Day and night images combined├── val/│ ├── day/│ │ └── 0/ ... 8/│ └── night/│ └── 0/ ... 8/├── test/│ ├── day/│ │ └── 0/ ... 8/│ └── night/│ └── 0/ ... 8/├── AOD_test/│ ├── 0.0-0.1/ ... 0.2-0.3/ ... # AOD bins (e.g., [AOD_range]/oktas)│ │ └── 0/ ... 8/└── adaptation_sets/ ├── moonless_train_data/ # Few-shot domain adaptation: moonless night │ └── 0/ ... 8/ └── aod_train_data/ # Few-shot domain adaptation: high AOD └── 0/ ... 8/</code></pre> Dataset key features Balanced Day/Night Ratio: All main subsets (Train, Validation, Test) maintain a 50/50 distribution of daytime and nighttime images. Balanced number of images per okta-class. Bias Reduction: The Test set (Germany) was cross-reviewed by a second independent observer to ensure ground-truth reliability. Low Temporal Redundancy: Consecutive images were excluded unless a distinct change in cloud cover occurred to minimize autocorrelation. Includes images from systems with triband filters (Spain) and standard IR-cut filters (Germany/La Palma), testing model robustness to small spectral shifts. All samples are 8-bit RGB images generated from High Dynamic Range (HDR) multi-exposure sequences. Usage and Citation This dataset is intended for research in meteorology, atmospheric sciences, and computer vision. It is part of the work: Longarela et al. (2026). Users should cite the following references regarding the GOA-SCAN network and the OMEA systems: Roman et al. (2025) for network and spectral response details. González-Fernández et al. (2024a) and González-Fernández et al. (2024b) for HDR processing and capture software