DCLS-GAN:利用生成对抗网络的天绘一号卫星高原地区影像去云方法

DCLS-GAN:cloud removal method for plateau area of TH-1 satellite image

利用深度学习开展高原地区卫星影像去云是一个研究热点。本文提出了基于DCLS-GAN的天绘一号卫星高原地区影像的去云方法,采用对抗学习的思想构建深度卷积对抗生成网络,自主学习影像中云覆盖部分的典型地表特征,从而恢复云覆盖下垫面形貌。基于Encoder-Decoder结构生成网络,构建固定与可移动2种云区掩膜,在矩形固定中心掩模预训练之后进行随机位置云掩模迁移训练,使用最小二乘重建损失与交叉熵对抗损失的联合损失函数,用于精确修复云覆盖区域地表;基于CNN鉴别网络,判别生成影像的真实性。采用双线性插值提高云覆盖区域的修复精度,后处理使用泊松编辑处理平滑预测边界,减少伪迹的影响。在测试数据集上的试验结果表明,本文方法的总体去云效果在峰值信噪比、结构相似性与自然影像无参考质量评价算法指标上优于经典方法与原始Context Encoder,速度上较经典图像重建方法优势较大,具有较好的实际应用前景。

It has been a research hotspot to apply deep learning to remove cloud on satellite images.In this paper,we propose a cloud removal method based on DCLS-GAN for the plateau image of TH-1 satellite.The generator is constructed with the structure of encoder-decoder,and two types of fixed and removable cloud masks are used in training.The least squarereconstruction loss and cross-entropy adversarial loss are used to generate the prediction image of cloud coverage area,whilel east square loss is also used in the discriminator to identify the authenticity of the generated image.Joint optimization of generator and discriminator is achieved by continuous iteration,after which,bilinear interpolation is used to improve the restoration accuracy of cloud coverage area,and Poisson editing is used to smooth the prediction boundary and reduce the influence of artifacts.The experimental results on the testing dataset show that the cloud removal effect of proposed method exceeds classical methods and the original Context Encoder in peak signal-to-noise ratio and structure similarity,and experiments on images with real cloud area also show that proposed method has lower indicators under blind image quality assessment.Finally,the speed is faster than classical methods and equals Context Encoder,thus it has a better practical application prospect.