Ground-based Visible-light Cloud Image Classification based on a Convolutional Neural Network

The classification of ground-based cloud maps is an important part of meteorological research and has been researched by many people over the years. Thus far, cloud classification has not been well solved in the fields of meteorology and image processing because of the complexity of cloud cover changes. Many researchers try to find more distinctive visual representations to describe different kinds of clouds, but most existing methods only use manual visual descriptors, and the results are not satisfactory. Inspired by the great progress of deep learning in large-scale image classification tasks, a deep convolutional neural network (CNN) structure that is trained from scratch is proposed to solve the relatively small-scale cloud classification problem. In two real data sets, the CNN shows a strong learning ability and classification effect, and the performance of the CNN is better than that of several previously proposed machine learning models that have been trained by manual texture descriptors. Through image processing, the mechanism used by convolutional neural networks when processing cloud images is analysed. This analysis provides theoretical support for further research involving processing more classifications and more complex foundational visible-light cloud images and explores the use of convolutional neural networks to classify ground-based cloud images.