Application of deep learning in cloud cover prediction using geostationary satellite images

Predicting cloud cover is essential in fields, such as agriculture, climatology, and meteorology, where accurate weather forecasting can significantly impact decision-making. Traditional methods for cloud cover prediction encounter significant limitations in capturing complete spatial and temporal cloud dynamics. To address the problem, this study employs high-resolution data from the new-generation geostationary satellite Geostationary Korea Multi-Purpose Satellite 2A (GEO-KOMPSAT-2A; GK2A) and enables more accurate and timely predictions when combined with advanced deep learning techniques. We explore the effectiveness of advanced deep learning techniques - specifically 3D Convolutional Neural Networks, Long Short-Term Memory networks, and Convolutional Long Short-Term Memory (ConvLSTM) - using GK2A cloud detection data, which provides updates every 10 minutes at 2 km spatial resolution. Our model utilizes training sequences of four past hourly images to predict cloud cover up to 4 hours ahead. For improved computational efficiency, each image is divided into four patches during training. Notably, this research incorporates a dynamic learning model, continuously updating with the most recent data, in contrast with static models which do not update their parameters with new data once trained. Results show that ConvLSTM tends to exhibit stable and relatively higher performance across prediction times compared to the other models in August 2021. While transformer models, such as Video Swin Transformer and TimeSformer, showed strong training performance, they struggled with overfitting, particularly on smaller datasets. In contrast, ConvLSTM demonstrated better generalization to test data, highlighting its suitability for tasks with limited training data and simpler structures. Year-long validation demonstrates the robustness of the ConvLSTM model, which consistently outperforms the other models in all major metrics, including a precision of 0.79, recall of 0.80, F1-score of 0.80 (which balances precision and recall), and accuracy of 0.78 throughout 2021. However, results show that the model's performance in terms of F1-score, recall, and precision is positively correlated with cloud fraction, with a slight tendency for higher accuracy during the summer compared to winter, indicating sensitivity to seasonal cloud cover variations.