Convective cloud identification in all-sky ground-based and satellite-based images using vision transformers

The detection of dangerous convective cloud types, specifically cumulonimbus and towering cumulus, is crucial for aviation safety and meteorological forecasting. Traditionally reliant on human observers, this process is limited by subjectivity, high costs, and lack of scalability. Although deep learning models, particularly convolutional neural networks (CNNs), have been employed for automated cloud analysis, relatively few studies have addressed hazardous convective cloud classification from ground-based sky imagery, and vision transformer architectures have yet to be thoroughly evaluated for this purpose. This research addresses that gap by assessing several transformer-based vision architectures on both daytime all-sky camera images and satellite imagery. The results show that transformer-based vision models outperform CNN baselines for both modalities, and achieving a Probability of Detection above 90% and False Alarm Ratios below 20% for ground-based camera imagery, thus meeting operational thresholds defined by meteorological authorities. These findings demonstrate the potential of using vision transformers in improving automated weather monitoring systems, thereby contributing to enhanced operation safety in aviation.