Automatic Cloud Segmentation Based on Fused Fully Convolutional Networks

Abstract

Cloud detection and segmentation of remote sensing images is a pivotal task in the area of weather forecast. Many meteorologic applications such as precipitation forecast, extreme weather forecast, etc., depend on the results of the cloud detection. In this paper, based on the satellite remote sensing image dataset, we propose an image segmentation model to address the cloud detection problem. Our model is derived from the fully convolutional neural network, which achieves pixel-level cloud segmentation results on high resolution, large scale, multi-channel satellite images. We introduce Deep Feature Aggregation and Model Fuse strategies to improve the cloud segmentation results. Compared with the traditional methods, our proposed algorithm has the advantages that is independent of the expert knowledge, totally data motivated, and more robust in hard cases. The testing results show that the proposed model can satisfy the requirements of the weather forecast, thus has a strong potential to be put into business usage.