SAR影像与对抗学习融合下的可见光影像云雾去除

Cloud removal of visible images utilizing the fusion of SAR imagery and adversarial learning

针对光学遥感影像经常受到云雾遮挡、干扰解译,使得数据利用率低的问题,该文利用SAR影像全天时全天候的特性,提出了一套“雷达-光学”的影像转换算法,以SAR影像为输入,利用对抗网络结合深度学习强大的非线性映射能力以及对抗学习博弈式学习任意数据分布的特点,实现通过云覆盖区域的SAR影像直接生成相应的无云伪可见光影像。实验表明,该方法对于薄云、厚云、大范围云雾的去除效果均比较好,能够在很大程度上提升光学遥感影像的可用性。

In response to the problem of low data utilization caused by frequent cloud and fog occlusion and interference interpretation in optical remote sensing images,this paper proposes a"radar-optics"image conversion algorithm based on the all-weather and all-weather characteristics of SAR images.Taking SAR images as input,the algorithm utilizes the strong nonlinear mapping ability of adversarial networks combined with deep learning,as well as the characteristics of adversarial learning game based learning for arbitrary data distribution,to directly generate corresponding cloud free pseudo visible light images through SAR images in cloud covered areas.The experiment shows that this method has good removal effects on thin clouds,thick clouds,and large-scale clouds and mist,and can greatly improve the availability of optical remote sensing images.