基于条件生成对抗模型的遥感卫星影像重建技术研究

为了快速侦察未知区域的地貌信息，遥感卫星可对特定区域进行扫描以获取遥感卫星影像。当卫星经过国外未知区域时，部分卫星无法针对某特定区域进行长时间的驻留扫描，本文提出一种基于条件生成对抗网络模型（Conditional Generative Adversarial Network，CGAN）进行网络训练，前期将某方法获取的区域轮廓地形信息作为CGAN网络的生成网络和鉴别网络中的条件约束信息，通过网络生成器与判别器在训练过程中互相博弈产生特定的输出集，有效地实现由单张电子轮廓图像到对应卫星遥感图像的端到端的非线性映射。本文通过原真实卫星遥感图像与生成卫星遥感图像进行四种对比误差计算，平均误差、均方误差与结构相似度均高于99%，峰值信噪比高于30 dB，生成的图像与原图像之间具备高相似度，实现了在获取坐标定位轮廓信息的先验条件下，对特定区域进行遥感卫星影像内容重建技术。

Research on remote sensing satellite image reconstruction technology based on conditional generation confrontation model

In order to quickly detect the geomorphic information in unknown areas, remote sensing satellites can scan specific areas to obtain remote sensing satellite images. When satellites pass through foreign areas, some satellites are unable to perform long-term resident scanning for a specific area. This paper proposes a Conditional Generative Adversarial Network (CGAN) for network training. In the early stage, the terrain information obtained by a certain method is used as the conditional constraint information in the CGAN network''s generation network and identification network, through the mutual game between the network generator and discriminator during the training process, a specific output set is generated, effectively achieve end-to-end nonlinear mapping from a single electronic contour image to the corresponding satellite remote sensing image. This article compares four types of error calculations between the original real satellite remote sensing image and the generated satellite remote sensing image. The average error, mean square error, and structural similarity SSIM are all higher than 99%, and the peak signal-to-noise ratio is higher than 30 dB. The generated image has high similarity with the original image, achieve the reconstruction technology of remote sensing satellite image content in specific areas under the prior condition of obtaining coordinate positioning contour information.