Artificial neural networks in automatic image classifications of cloud from ground‐based observations using deep learning models

Cloud classification is a critical task in meteorology, with applications in weather forecasting, climate modelling, and environmental monitoring. Traditionally, cloud observations are made visually by experienced observers, which can introduce human errors and inconsistencies. This study aims to develop accurate deep learning models for automated cloud classification from ground‐based images. A dataset of over 200,000 cloud photographs was collected from professional observers at meteorological stations of the Institute of Meteorology and Water Management–National Research Institute. The images were preprocessed and annotated with cloud type labels. Convolutional neural networks and transformer models were employed for feature extraction, combined with multilayer perceptrons for classification. Different architectures were explored, including ResNet, EfficientNet, Vision Transformers, and their variants. Hyperparameter tuning and data augmentation techniques like RandAugment were applied to improve generalisation. The best‐performing model achieved 97.4% accuracy for classifying four common cloud genera: cirrus, cumulus, stratus, and clear sky. Additional experiments classified up to 11 genera, with accuracy decreasing as complexity increased, due to data limitations. Error analysis revealed confusions between visually similar classes. While promising, limitations exist from using partial cloud views versus whole‐sky imagery. We start with four classes and end with 11 classes, progressively showing subsequent errors in the cloud genera classification process, something not presented in other publications. Future work involves collecting a more balanced dataset with standardised protocols. Integration of all‐sky cameras could help address current restrictions. This research demonstrates the viability of deep learning for automated cloud observation, with opportunities to advance meteorological applications through continued methodological refinement.