Semantic Segmentation of Clouds in Satellite Imagery Using Deep Pre-trained U-Nets

Abstract

Earth observation and remote sensing technologies are widely used in various application areas. Because the abundance of collected data requires automated analytics, many communities are utilizing deep convolutional neural networks for such tasks. Automating cloud detection in remote sensing and earth observation imagery is a useful prerequisite for providing quality imagery for further analysis. In this paper, we train a model that uses a deep convolutional U-Net architecture, utilizing transfer learning to perform semantic segmentation of clouds in satellite imagery. Our proposed model outperforms state-of-the-art networks on a benchmark dataset based on several relevant segmentation metrics, including Jaccard Index (+7.69%), precision (+6.21%), and specificity (+0.37%). Moreover, we demonstrate that transfer learning utilizing a 4-channel input into a U-Net architecture is possible and highly performant by using a deep ResNet-style architecture pre-trained on ImageNet for the initialization of weights in three channels (red, green, and blue bands) and random initialization of weights in the fourth channel (near infrared band) of the first convolutional layer of the network.