基于稠密残差网络的多序列卫星图像去云

Multi-sequence satellite image cloud removal based on dense residual network

遥感影像中最常见的问题是云层污染,它会导致图像信息缺失,降低遥感数据的可用性。针对该问题,提出了一种基于稠密残差网络的多序列卫星图像去云算法。首先,该网络使用多序列的有云卫星图像作为输入,能为网络提供更多的时序特征信息,提升去云效果;其次,在网络中段使用稠密残差层,以保证卷积层之间最大程度地传递和使用特征信息,让生成的修复图像整体结构合理、边缘细节更加清晰;最后,使用像素上采样来增强空间信息的利用,提升修复效果。该方法在欧洲“哨兵-2”遥感卫星图像数据集上进行验证,峰值信噪比和结构相似度指标为27.59和0.8540,两项指标均超过了该数据集的原处理方法STGAN,提升了遥感图像去云的效果。

The most common problem in remote sensing imagery is cloud pollution,which will lead to lack of image information and low availability of remote sensing data.To solve this problem,this paper proposed a cloud removal algorithm for multi-sequence satellite images based on dense residual network.The network used multi-sequence cloud satellite images as input,which could provide more timing feature information for the network and improve the effect of cloud removal.Secondly,using a dense residual layer in the middle of the network,which ensured the maximum transfer and use of feature information between the convolutional layers,so that the generated repair image as a whole structure was reasonable and the edge details were clearer.Finally,it used pixel shuffle upsample to enhance the use of spatial information and improved the repair effect.This method was verified on the European“Sentinel-2”remote sensing satellite image dataset.The peak signal-to-noise ratio and structural similarity index are 27.59 and 0.8540,both of which exceed the original processing method STGAN of the data set,and improve the effectiveness of remote sensing image to cloud.