Automatic cloud classification from all-sky images is critical for meteorological monitoring, yet most research focuses on neural network architecture, ignoring data quality. This paper compares 23 convolutional neural network (CNN) configurations using the TJNU GRSCD dataset, which is appropriate for the specific hardware requirements of the task. The results showed that all architectures achieve an accuracy limit of 85–88%. The analysis revealed a concentration of errors in visually overlapping classes, indicating a fundamental inconsistency (or subjectivity) in the labeling rather than model deficiencies. It is concluded that a data-centric approach (strict dataset verification by experts) is necessary before automated processing of an unlabeled archive of approximately 2 million panoramic images.
all-sky camera, cloud classification, convolutional neural networks, comparative testing of architectures, subjectivity in data labeling, data quality
1. Dokumentaciya Ultralytics YOLO [Elektronnyy resurs]. – Rezhim dostupa: https://docs.ultralytics.com/models (data obrascheniya: 27.03.2026).
2. Tian Y., Ye Q., Doermann D. Yolov12: Attention-centric real-time object detectors //arXiv preprint arXiv:2502.12524. – 2025.
3. Tan M., Le Q. Efficientnet: Rethinking model scaling for convolutional neural networks //International conference on machine learning. – PMLR, 2019. – S. 6105-6114.
4. He K. et al. Deep residual learning for image recognition //Proceedings of the IEEE conference on computer vision and pattern recognition. – 2016. – S. 770-778.
5. HuaYun Ground Based Cloud Image Dataset [Elektronnyy resurs]. – Rezhim dostupa: https://github.com/SadaharuZL/HuaYun-BJUT-MIP-Cloud-Dataset (data obrascheniya: 27.03.2026).
6. Cirrus Cumulus Stratus Nimbus (CCSN) Database [Elektronnyy resurs]. – Rezhim dostupa: https://www.kaggle.com/datasets/mmichelli/cirrus-cumulus-stratus-nimbus-ccsn-database (data obrascheniya: 27.03.2026).
7. Ground-based cloud images (SWIMCAT) [Elektronnyy resurs]. – Rezhim dostupa: https://data.mendeley.com/datasets/vwdd9grvdp/2 (data obrascheniya: 27.03.2026).
8. (Clouds-1000 [Elektronnyy resurs]. – Rezhim dostupa: https://data.mendeley.com/datasets/4pw8vfsnpx/2 (data obrascheniya: 27.03.2026).
9. TJNU Ground-based Remote Sensing Cloud Database [Elektronnyy resurs]. – Rezhim dostupa: https://github.com/shuangliutjnu/TJNU-Ground-based-Remote-Sensing-Cloud-Database (data obrascheniya: 27.03.2026).
10. Liu S. et al. Ground‐based cloud classification using task‐based graph convolutional network //Geophysical Research Letters. – 2020. – T. 47. – №. 5. – S. e2020GL087338.
11. Galileyskiy V.P., Elizarov A.I., Kokarev D.V., Matvienko G.G., Morozov A.M. Programmno-tehnicheskaya sistema monitoringa i kontrolya optiko-fizicheskogo sostoyaniya atmosfery. Chast' 1. Panoramno-opticheskaya stanciya // Optika atmosfery i okeana. Fizika atmosfery: Sbornik dokladov XXVII Mezhdunarodnogo simpoziuma. [Elektronnyy resurs]. ISBN 978-5-94458-187-7. - Tomsk: Izd-vo IOA SO RAN, 2021. C. F5-F9.
12. Hamid O. H. From model-centric to data-centric AI: A paradigm shift or rather a complementary approach? //2022 8th International Conference on Information Technology Trends (ITT). – IEEE, 2022. – S. 196-199.
13. Sambasivan N. et al. “Everyone wants to do the model work, not the data work”: Data Cascades in High-Stakes AI //proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. – 2021. – S. 1-15.
