Using Deep Learning for Enhancing the Performance of Ground-based Cloud Images Classification

Eugen MIHULEȚ, Gabriela CZIBULA, Ștefan ALEXANDRESCU, Ana-Maria MARDALOESCU, Alexandra-Ioana ALBU, Mariana-Ioana MAIER

Abstract

Clouds are important for Earth’s climate, as they are the source of precipitation, they control the amount of solar energy that reaches the surface of this planet, and they also influence the Earth`s Radiation Budget. Thus, cloud classification is important as it helps in the context of weather forecasting and for supervising climate changes. Clouds are organised in many forms, their composition and density are variable and their colour and height in the sky can differ, which makes their classification a challenging one. This paper proposes two Xception-based convolutional neural network architectures, namely X-Cloud and M-Cloud, for classifying ground-based cloud images and analyses their prediction performance and their ability of learning to extract and classify the features of distinct cloud types. Experiments were performed on two ground-based cloud data sets of different sizes and with different characteristics, with and without using data augmentation techniques for improving the training data sets. The performance of the best proposed model, namely X-Cloud, is compared with that of two other deep learning architectures employed in the literature for cloud classification. The comparative analysis highlights an improvement of about 10%, on average, in the case of the X-Cloud model, in terms of the F-measure performance metric.

Keywords

Cloud classification, Computer vision, Deep learning.

View full article