结合空洞卷积的条件对抗网络遥感影像生成  

Conditional adversarial network remote sensing image generation combined with atrous convolution

作　　者：杨汝春 李佳田[1] 吕少云 张兴忆 张泽龙 YANG Ruchun;LI Jiatian;LYU Shaoyun;ZHANG Xingyi;ZHANG Zelong(Faculty of Land Resource Engineering,Kunming University of Science and Technology,Kunming 650093,China)

机构地区：[1]昆明理工大学国土资源工程学院,昆明650093

出　　处：《测绘科学》2022年第6期90-98,118,共10页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(41561082)

摘　　要：针对利用土地覆盖与数字高程模型生成遥感影像存在部分细节缺失的问题,该文提出结合空洞卷积的条件对抗网络以生成高分辨率遥感影像。首先,生成器网络采用空间自适应归一化结构将土地覆盖特征与数字高程模型特征融合;其次,将并行空洞卷积模块嵌入生成器网络中的编码器中,以改善融合后的特征;最后,解码器输出的影像输入判别器网络进行判别,输出生成结果。为验证该文网络的有效性,与现有的条件生成对抗网络在GeoNRW和DFC2020数据集上进行实验。结果表明:该文网络在GeoNRW数据集上的距离得分较对比网络平均降低0.11,整体平均交并比较对比网络平均提升0.77%,生成的影像地物细节更丰富;此外,该文网络在DFC2020数据集上具有一定的泛化能力,生成的影像优于对比网络生成的影像。In view of the lack of details in remote sensing images generated by land cover and digital elevation model,a conditional adversarial network combined with atrous convolution was proposed to generate high-resolution remote sensing images.Firstly,the generator network used the spatially adaptive normalization structure to fuse the land cover features with the digital elevation model features.Secondly,the parallel atrous convolution module was embedded in the encoder in the generator network to improve the fused features.Finally,the decoder output images were inputed the discriminator network performs discrimination and outputed the generated result.In order to verify the effectiveness of the network in this paper,experiments with the existing conditional generation confrontation network were conducted on the GeoNRW and DFC2020 datasets.The results showed that the frechet inception distance score of this network on the GeoNRW dataset was 0.11 lower than that of the comparison networks on average,and the overall mean intersection over union was increased by 0.77%on average compared with the comparison networks.The generated image features had more details;in addition,the network of this paper had a certain degree of generality on the DFC2020 dataset.The image generated was better than the image generated by the comparison networks.

关 键 词：土地覆盖 数字高程模型 遥感影像 条件对抗网络 空洞卷积 

分 类 号：P237[天文地球—摄影测量与遥感]