Understanding Cloud Structures from Satellite Images using Deep Learning

With the advancement of remote sensing technology, many high-resolution satellite images have been produced that can be used in a variety of disciplines, including assessing crop yields and monitoring land cover and pollution in the ocean and on land. However, the hit back to this occurs when thick clouds block the path of identification of ground level objects. Around 65% of the earth’s outer surface is covered with clouds hence this has become a major problem. To encounter this issue, we discuss some of the deep learning techniques like convolution neural networks (CNN), for the detection of cloud shapes. The semantic segmentation technique namely U-Net was chosen in this paper as the base architecture as it works well on very fewer training samples and provides a better result. Also, the computation cost is very less which is a major advantage compared to other segmentation models as the non-availability of freely available data owing to many security and investment issues. The paper introduces a new architecture for this problem statement which is an improvement to the U-Net architecture which majorly uses the concept of altering the skip connections in different to improve the way the information is passed on. The dataset used here is a derived version of NASA Worldwide view data which is widely used for cloud segmentation. The evaluation results show significant improvement in terms of accuracy.