Weakly-supervised cloud detection and effective cloud removal for remote sensing images

We propose a one-stop automatic cloud processing scheme for RSI, including weakly supervised cloud detection and effective cloud removal. First, to avoid using massive expensive paired-labeled data, we use the idea of adversarial training to detect clouds, that is, randomly selecting some unpaired cloud and cloudless images to alternatively train a spectral-normalized Markov discriminator to obtain cloud masks. Second, a two-stage downsampling-restoring-upsampling-refining scheme is used to remove detected clouds. In order to further improve the rationality and fineness, we construct a fractional-order anisotropic filter kernel and design a convolution process to impose spatial regularization constraints on the loss construction at each stage, taking to account confidence values and structural priorities. Comprehensive experimental results show that our proposed networks are superior to the SOTA methods in terms of objective indicators and subjective performance. In addition, the performance of our one-stop approach in automatically detecting and eliminating clouds is also very satisfactory.