DeepSky dataset: A new benchmark for ground-based cloud classification using all-sky images

Cloud observation methods have attracted growing interest due to their significant role in climate prediction and weather forecasting applications since clouds serve as important indicators of weather conditions. Ground-based cloud image classification, which involves categorizing all-sky images into predefined classes, is a key approach to address this challenge. Deep learning methods have demonstrated their effectiveness in various computer vision tasks; however, their performance is inherently tied to the size and quality of the dataset employed for training and evaluation. Many ground-based all-sky image datasets are available, each fulfilling various criteria, including the number of samples, the duration of data collection, the interval between collected samples, the geographical location of data collection, the resolution of images, the types of sky images, and the annotation process used to assign labels to the images In this paper, we introduce the DeepSky dataset, a novel ground-based all-sky image classification dataset captured over a span of two years. The dataset comprises over seven thousand images and includes seven cloud categories, making it one of the largest datasets in the field. Additional to the dataset, we present an effective cloud classification model based on SWIN Transformers, which achieves state-of-the-art accuracy when evaluated to the other large all-sky image dataset. We are confident that the new dataset, the analysis on cross-dataset evaluation, and the transformer-based proposed method will provide researchers with new opportunities and challenges to explore and expand upon in their investigations. The dataset and supplementary code materials can be accessed from: https://github.com/dimkastan/DeepSky-classification-dataset.