Sea ice SAR-to-Optical image translation based on improved CGAN
Remote sensing sea ice monitoring is a current research hotspot.By translating sea ice SAR images into optical images through the Conditional Generative Adversarial Network(CGAN),all-weather and intuitive monitoring data can be obtained.However,the translation results obtained by this method have problems such as blurring of contours,weakening of textures,and color distortion.This article designs an improved CGAN network to address the above issues.Taking into account the current improvement methods,the new model adds a hollow space pyramid module to the network structure and designs a skip connection with a cross feature fusion module.The loss function uses a joint loss function of structural similarity and L1 norm.This article selects 5 Sentinel-1 images and 7 Sentinel-2 images from the East Beaufort Sea area for experiments.The experimental results show that the improved CGAN translated images have better visual effects,with peak signal-to-noise ratio(PSNR)increased by 3.4,structural similarity(SSIM)increased by 0.11,root mean square error(RMSE)decreased by 13%,and the accuracy of the processed images for sea ice classification improved by 7.33%compared to SAR images.
NETL
NSTL
万方数据
