Abstract: Remote sensing image data annotation is time-consuming and costly and requires expert knowledge, making it difficult to obtain remote sensing data with labels. Therefore, it is necessary to generate an effective method of remote sensing data with labels. Starting from the cycle-consistent generative adversarial networks for style transfer in the field of computer vision, a simulated remote sensing image conversion method based on deep learning and cycle-consistent generative adversarial networks to generate new dataset is proposed. This method regards the source data and the generated data as the source domain and the target domain, which can be regarded as the style transfer of the simulated remote sensing dataset. The generated dataset can be further used for common tasks of remote sensing images, such as classification, semantic segmentation, and domain adaptation. Experimental results show that this method can effectively generate simulated remote sensing data with style transfer.

Key words: simulated remote sensing, style transfer, cycle-consistent generative adversarial network, domain adaptation