A NEW CLOUD FEATURE DIFFERENCE LOSS FOR ENHANCING THE DETECTION OF CLOUDS IN REMOTE SENSING IMAGES

Research has shown that most of the Earth's surface is covered by clouds, which have reduced the usability of optical remote sensing images. Therefore, it is critical to detect clouds quickly and accurately. Cloud detection methods based on deep learning have been widely studied in recent years. However, the existing methods still face a significant challenge that thin clouds are often translucent and easily missed. To enhance the detection of thin clouds by convolutional neural networks, we propose a cloud feature difference (CFD) loss, which gathers the samples in thin clouds, thick clouds and subsurface as a sample group, namely Cloud-Triplet. By measuring the features difference between samples, the CFD loss forces the network to model discriminative features during training, thus improving the ability of detecting thin clouds. Experiments show that our proposed CFD loss is effective in enhancing the detection of clouds.