Cloud-type Classification of Ground-Based Images using Deep Learning

Precise cloud detection and classification is a crucial element for various meteorological applications, such as solar irradiance forecast or cloud cover estimation. In this paper, we present an automatic method for cloud-type classification from ground-based images using improved pretrained deep neural network. In the past, there were used algorithms using mainly low-level features. Our approach is to use artificial neural network (ANN), namely convolutional neural network (CNN) because these type of networks use both types of image features: the first couple convolutional layers learn filters for finding lines, dots, curves, etc. while the later layers learn to recognize common objects and shapes. Therefore, with this approach we are able to reach high precision of cloud-type classification with overall accuracy of 85%.