Vol. 39, issue 11, article # 9

Elizarov A. I., Shaleev A. V. Semantic ambiguity of training samples as a limiting factor of accuracy of convolutional networks in cloud classification. // Optika Atmosfery i Okeana. 2026. V. 39. No. 11. P. 986–994. DOI: 10.15372/AOO20261109 [in Russian].
Copy the reference to clipboard

Abstract:

Automatic cloud classification from all-sky images plays a key role in meteorological monitoring. However, most research focuses on finding the optimal neural network architecture, ignoring the quality of the source data. This paper examines the impact of semantic ambiguity in reference markup on the performance of deep learning methods. To empirically prove that data quality is the main limiting factor of accuracy, a comprehensive test of 23 convolutional network architectures was conducted on the Tianjin Ground-based Remote Sensing Cloud Database (TJNU GRSCD) dataset. The results showed that, regardless of the network topology, a fundamental accuracy limit of 85–88% is achieved. The error analysis revealed their concentration on visually indistinguishable classes (mixed, altocumulus, cirrus), confirming the optical-physical and logical inconsistency of the training set. It is concluded that a transition to a data-quality-oriented approach and rigorous expert verification of training sets taking into account multimodal meteorological data is necessary before scaling the algorithms. The results of this work lay the foundation for the automated processing of an unlabeled archive containing approximately 2 million all-sky images from two active monitoring stations.

Keywords:

all-sky camera, cloud classification, convolutional neural network, comparative testing of architecture, semantic ambiguity, data quality

Figures:

References:

1. Urquhart B., Kurtz B., Dahlin E., Ghonima M., Shields J.E., Kleiss J. Development of a sky imaging system for short-term solar power forecasting // Atmos. Meas. Tech. 2015. V. 8, N 2. P. 875–890. DOI: 10.5194/amt-8-875-2015.
2. Bauer V. Design and Evaluation of an Affordable Open-Source All-Sky Camera for Measuring Diffuse Sky Radiance: Diploma Thesis. Wien: Technische Universität Wien, 2025. 115 p. DOI: 10.34726/hss.2025.120149.
3. Galileiskii V.P., Elizarov A.I., Kokarev D.V., Matvienko G.G., Morozov A.M. Software and hardware system for monitoring and control of the optical-physical state of the atmosphere: Part 1: Panoramic optical station // Proc. SPIE. 2021. V. 11916. P. 1798–1801. DOI: 10.1117/12.2603231.
4. Abramova I.A., Demchev D.M., Kharyutkina E.V., Savenkova E.N., Sudakow I.A. Primenenie svertochnoi neironnoi seti U-Net i ee modifikatsii dlya segmentatsii tundrovykh ozer na sputnikovykh opticheskikh izobrazheniyakh // Optika atmosf. i okeana. 2024. V. 37, N 1. P. 48–53. DOI: 10.15372/AOO20240106; Abramova I.A., Demchev D.M., Kharyutkina E.V., Savenkova E.N., Sudakow I.A. Utilization of the U-Net convolutional neural network and its modifications for segmentation of Tundra Lakes in satellite optical images // Atmos. Ocean. Opt. 2024. V. 37, N 2. P. 205–210.
5. Bloshchinskiy V.D., Kramareva L.S., Shamilova Yu.A. Detektirovanie oblachnogo pokrova s ispol'zovaniem neironnoi seti po dannym pribora MSU-GS kosmicheskogo apparata «Arktika-M» N 1 // Optika atmosf. i okeana. 2024. V. 37, N 2. P. 99–104. DOI: 10.15372/AOO20240202; Bloshchinskiy V.D., Kramareva L.S., Shamilova Yu.A. Cloud cover detection using a neural network based on MSU-GS instrument data of Arktika-M No. 1 satellite // Atmos. Ocean. Opt. 2024. V. 37, N 3. P. 394–399.
6. Luo J., Pan Y., Su D., Zhong J., Wu L., Zhao W., Hu X., Qi Z., Lu D., Wang Y. Innovative cloud quantification: deep learning classification and finite-sector clustering for ground-based all-sky imaging // Atmos. Meas. Tech. 2024. V. 17, N 12. P. 3765–3781. DOI: 10.5194/amt-17-3765-2024.
7. Tian Y., Ye Q., Doermann D. Yolov12: Attention-centric real-time object detectors // arXiv. 2025. URL: https://arxiv.org/abs/2502.12524. DOI: 10.48550/ arXiv.2502.12524.
8. He K., Zhang X., Ren S., Sun J. Deep residual learning for image recognition // Proc. IEEE. 2016. P. 770–778. DOI: 10.1109/CVPR.2016.90.
9. Liu S., Duan L., Zhang Z., Cao X., Durrani T.S. Ground-based remote sensing cloud classification via context graph attention network // IEEE Trans. Geosci. Remote Sens. 2021. V. 60. P. 1–11. DOI: 10.1109/ TGRS.2021.3063255.
10. Longarela Fuente B., Román R., Buitrago F., González-Fernández D., Gatón J., Herrero Del Barrio C., Herrero-Anta S., González R., Mateos D., Doppler L., Barreto Á., Toledano C., Cachorro V.E. Cloud Cover Retrieval from All-sky Images Based on Neural Networks. 2026. P. 1–33. DOI: 10.2139/ssrn.6560840.
11. HuaYun-BJVT-MIP-Ground Based Cloud-Dataset. URL: https://github.com/SadaharuZL/HuaYun-BJUT- MIP-Cloud-Dataset (last access: 27.03.2026).
12. Cirrus Cumulus Stratus Nimbus (CCSN) Database. URL: https://www.kaggle.com/datasets/mmichelli/ cirrus-cumulus-stratus-nimbus-ccsn-database (last access: 27.03.2026).
13. Ground-based Cloud Images (SWIMCAT). URL: https: //data.mendeley.com/datasets/vwdd9grvdp/2 (last access: 27.03.2026).
14. Clouds-1000. URL: https:// data.mendeley.com/datasets/4pw8vfsnpx/2 (last access: 27.03.2026).
15. TJNU Ground-based Remote Sensing Cloud Database. URL: https://github.com/shuangliutjnu/TJNU-Ground- based-Remote-Sensing-Cloud-Database (last access: 27.03.2026).
16. Liu S., Li M., Zhang Z., Cao X., Durrani T.S. Ground-based cloud classification using task-based graph convolutional network // Geophys. Res. Lett. 2020. V. 47, N 5. P. 1–8. DOI: 10.1029/2020GL087338.
17. Zha D., Bhat Z.P., Lai K.H., Yang F., Jiang Z., Zhong S., Hu X. Data-centric artificial intelligence: A survey // ACM Comput. Surv. 2025. V. 57, N 5. P. 1–42. DOI: 10.1145/3711118.
18. Hamid O.H. From model-centric to data-centric AI: A paradigm shift or rather a complementary approach? // 8th International Conference on Information Technology Trends (ITT). 2022. P. 196–199. DOI: 10.1109/ ITT56123.2022.9863935.
19. Sambasivan N., Kapania S., Highfill H., Akrong D., Paritosh P., Lora M. “Everyone wants to do the model work, not the data work”: Data Cascades in High-Stakes AI // Proc. CHI Conference on Human Factors in Computing Systems. 2021. N 39. P. 1–15. DOI: 10.1145/3411764.3445518.
20. Liu Z., Lin Y., Cao Y., Hu H., Wei Y., Zhang Z., Lin S., Guo B. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows // Proc. IEEE/CVF International Conference on Computer Vision (ICCV). 2021. P. 10012–10022.
21. Liu S., Yu Z., Zhang Z., Shi C., Xiao B. Hierarchical Fusion Transformer for Multimodal Ground-Based Cloud Type Classification // IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 2025. V. 18. P. 25192–25203. DOI: 10.1109/JSTARS.2025.3614756.
22. Simonyan K., Zisserman A. Very deep convolutional networks for large-scale image recognition // arXiv. 2014. URL: https://arxiv.org/abs/1409.1556. DOI: 10.48550/arXiv.1409.1556.
23. Szegedy C., Vanhoucke V., Ioffe S., Shlens J., Wojna Z. Rethinking the inception architecture for computer vision // Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016. P. 2818–2826. DOI: 10.1109/CVPR.2016.308.
24. Howard A., Sandler M., Chu G., Chen L.-C., Chen B., Tan M., Wang W., Zhu Y., Pang R., Vasudevan V., Le Q.V., Adam H. Searching for MobileNetV3 // Proc. IEEE/CVF International Conference on Computer Vision (ICCV). 2019. P. 1314–1324. DOI: 10. 1109/ICCV.2019.00140.
25. Li S., Wang M., Sun S., Wu J., Zhuang Z. CloudDenseNet: Lightweight ground-based cloud classification method for large-scale datasets based on reconstructed DenseNet // Sensors. 2023. V. 23, N 18. P. 1–25. DOI: 10.3390/s23187957.
26. Hu F., Hou B., Zhu W., Zhu Y., Zhang Q. Cloudy-Net: A deep convolutional neural network architecture for joint segmentation and classification of ground-based cloud images // Atmosphere. 2023. V. 14, N 9. P. 1–17. DOI: 10.3390/atmos14091405.
27. Lv Q., Li Q., Chen K., Lu Y., Wang L. Classification of ground-based cloud images by contrastive self-supervised learning // Remote Sens. 2022. V. 14, N 22. P. 1–18. DOI: 10.3390/rs14225821.
28. Ultralytics DOCS. URL: https://docs.ultralytics.com/ (data obrashcheniya: 31.05.2026).
29. Liu S., Li M., Zhang Z., Xiao B., Durrani T.S. Multi-evidence and multi-modal fusion network for ground-based cloud recognition // Remote Sens. 2020. V. 12, N 3. P. 1–20. DOI: 10.3390/rs12030464.
30. Multimodal-Ground-based-Cloud-Database. URL: https: //github.com/shuangliutjnu/Multimodal-Ground-based- Cloud-Database (last access: 27.03.2026).