Research on Semantic Segmentation of Ground- Based Cloud Image Based on Fully Convolutional Network

The production, dissipation, and variation of cloud play important roles in the weather forecast such as rain, snow and hail, which are typically reflected in changes in cloud types and cloud cover. It is critical to obtain the cloud cover of each type of cloud quickly and accurately in the constantly changing ground-based cloud images in order to capture and predict weather changes, and the prerequisite is to implement automatic ground-based cloud image segmentation technology classified by cloud type. To implement this technique, this paper introduces a full convolutional network structure in deep learning to implement a segmentation algorithm for ground- based cloud images at the semantic level. We first create the ground-based cloud image database for semantic segmentation (GBCSS). This database includes ten categories of cloud and one category of background for a total of eleven categories, with pixel-level labels for each image. Then, using convolutional neural network structures, we design five full convolutional networks and evaluate the semantic segmentation effects on GBCSS. The experimental results show that all five full convolutional networks have good segmentation accuracy of on GBCSS, with the highest accuracy reaching 75.7%, laying the foundation for the application of semantic segmentation in ground-based cloud image.