基于 GAN 网络的电力设施遥感图像去云算法研究

针对电力设施遥感图像云层遮挡问题，提出了基于生成对抗网络(GAN) 的遥感图像去云算法。以条件生成对抗网络 (cGAN) 为主体结构，在编码器自适应填充卷积，设计了基于 Soft Attention的递归神经网络模块，通过对所有特征节点增加 全局依赖关系来解决网络局部最优问题，通过空间信息转换提取关键信息，提高去云与重建效果。实验结果表明，方法对遥 感图像中云层遮挡去除效果较好，重建图像的结构相似性(SSIM) 与峰值信噪比(PSNR) 分别达到0.983与32.899,分别提高 了23.93%与8.86%,均优于其他改进型GAN 网络。研究的方法不仅为基于遥感图像电力设施识别提供了基础，深度学习 遥感图像处理应用提供了参考。

Research on GAN based remote sensing image de-cloud algorithm for electric utilities

Aiming at occlusing cloud in remote sensing images of electric power facilities,a de-cloud model based on generative adversarial network(GAN)is proposed.Taking cGAN as the main structure,the model was designed with the encoder adaptive filling convolution,the recurrent neural network module based on soft attention,which is used to to solve the problem of network local optimum by adding global dependencies to all feature nodes.By extracting the key information through spatial information transformation,the cloud removal and reconstruction quality is improve.The experimental results demonstrated that the method has a better effect on cloud occlusion removal in remote sensing images.Compared with other GAN networks,the SSIM and PSNR of the reconstructed images reach 0.983 and 32.899, which are improved by 23.93%and 8.86%,respectively.The method not only provides a basis for remote sensing image-based power facility identification,but also provides a reference for deep learning remote sensing image processing applications,