Synthetic aperture radar (SAR) enables imaging of topographic surfaces day and night in different atmospheric conditions. SAR imaging systems record both intensity and phase information of the backscattered signals. Acquired intensity information is often exposed to speckle noise, and gathered phase information is usually corrupted by thermal and other types of noise. Thus, these types of noises have negative effects on interpretation of SAR images. Digital elevation model (DEM) can be generated by interferometric SAR using two SAR images, of the same area, with slightly different look angles. The generated DEM is affected by the corruption of both intensity and phase information. In this paper, a proposed framework of convolutional neural network (CNN) and modified Wiener filter (MWF) is suggested in DEM generation process. The main purpose of the proposed framework is minimizing not only speckle noise of input SAR images but also phase noise of the interferogram. Thus, an enhanced DEM can be generated. Extensive experiments are carried out and different DEMs are generated from original SAR and from both despeckled SAR images and filtered interferogram. Results and comparative analyses show significant improvements in both quality and vertical accuracy of the DEM generated by the proposed hybrid (CNN-MWF) framework.

合成孔径雷达 (SAR) 能够在不同的大气条件下昼夜对地形表面进行成像。SAR 成像系统记录反向散射信号的强度和相位信息。获取的强度信息通常会受到散斑噪声的影响，而收集的相位信息通常会被热噪声和其他类型的噪声破坏。因此，这些类型的噪声对 SAR 图像的解释具有负面影响。数字高程模型 (DEM) 可以通过干涉 SAR 生成，使用相同区域的两幅 SAR 图像，视角略有不同。生成的 DEM 受到强度和相位信息损坏的影响。在本文中，提出了在 DEM 生成过程中提出的卷积神经网络 (CNN) 和修改维纳滤波器 (MWF) 的框架。所提出框架的主要目的是不仅最小化输入 SAR 图像的散斑噪声，而且最小化干涉图的相位噪声。因此，可以生成增强的 DEM。进行了大量实验，并从原始 SAR 和去斑 SAR 图像和滤波干涉图生成不同的 DEM。结果和比较分析表明，所提出的混合 (CNN-MWF) 框架生成的 DEM 的质量和垂直精度都有显着提高。