Abstract:The article investigates deep learning architectures for automatic image classifications of cloud from ground‐based observations purpose. Artificial neural network models of varying complexities were evaluated using confusion matrices, while PCA and t‐SNE analyses revealed frequent misclassifications among cloud types. The research provides comprehensive error indications based on the number of cloud classes considered in models, highlighting the challenges in accurate cloud type differentiation. 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.

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

Cloud type classification using deep learning with cloud images

Cloud Identification from All-sky Camera Data with Machine Learning

Cloud Detection for Satellite Imagery Using Deep Learning

CloudDenseNet: Lightweight Ground-Based Cloud Classification Method for Large-Scale Datasets Based on Reconstructed DenseNet

Satellite Imagery-Based Cloud Classification Using Deep Learning

Machine learning of cloud types in satellite observations and climate models

Machine-learned cloud classes from satellite data for process-oriented climate model evaluation

Classification of Cloud Particle Imagery from Aircraft Platforms Using Convolutional Neural Networks

Cloud Detection Algorithm for Multi-Satellite Remote Sensing Imagery Based on a Spectral Library and 1D Convolutional Neural Network

AERIAL POINT CLOUD CLASSIFICATION WITH DEEP LEARNING AND MACHINE LEARNING ALGORITHMS

Sky Image Classification Based on Transfer Learning Approaches

Thin Cloud Removal for Multispectral Remote Sensing Images Using Convolutional Neural Networks Combined With an Imaging Model

Cloud classification through machine learning and global horizontal irradiance data analysis

Cloud Classification by Machine Learning for Geostationary Radiation Imager

Deep learning based cloud detection for medium and high resolution remote sensing images of different sensors

CloudY-Net: A Deep Convolutional Neural Network Architecture for Joint Segmentation and Classification of Ground-Based Cloud Images

Improved RepVGG ground-based cloud image classification with attention convolution

Cloud Classification with Unsupervised Deep Learning

Remote Sensing Image Cloud Detection Using a Shallow Convolutional Neural Network