Abstract
The traditional interferometric synthetic aperture radar denoising methods normally try to estimate the phase fringes directly from the noisy interferogram. Since the statistics of phase noise are more stable than the phase corresponding to complex terrain, it could be easier to estimate the phase noise. In this article, phase noises rather than phase fringes are estimated first, and then they are subtracted from the noisy interferometric phase for denoising. The denoising convolutional neural network is introduced to estimate the phase noise and then a modified network called IPDnCNN is constructed for the problem. Based on the IPDnCNN, a novel interferometric phase noise reduction algorithm is proposed, which can reduce the phase noise while protecting fringe edges and avoid the use of filter windows. The experimental results using the simulated and real data are provided to demonstrate the effectiveness of the proposed method.
Highlights
SYNTHETIC Aperture Radar Interferometry (InSAR) is an all-time and all-weather remote-sensing technique and can be used for generating digital elevation models (DEMs) or detecting surface deformation [1], [2]
It can be seen that the IPDnCNN method has reduced noise significantly while preserving the edge, whereas the slope adaptive filter and the improved Goldstein filter are less capable of denoising the interferometric phase
The proposed IPDnCNN is constructed based on denoising convolutional neural network (DnCNN)
Summary
SYNTHETIC Aperture Radar Interferometry (InSAR) is an all-time and all-weather remote-sensing technique and can be used for generating digital elevation models (DEMs) or detecting surface deformation [1], [2]. The Lee filter is designed to achieve a balance between residual noise and detail information loss [5], where a window with the adjustable size and direction is employed according to the local gradient of the interferogram This method only calculates 16 discrete orientations, which brings distortion to curved fringes. In [26], an adaptive multiresolution technique was proposed to modify the LFF estimation by setting a threshold to eliminate the “bad LFF values” which have a large difference compared to its neighboring pixels It provides better protection for phase fringes, but it is still hard to estimate the fringe frequency for highly sloped terrain. The number of samples used for noise training, 300000 here, is huge and all pixels of noisy interferogram are exploited in phase noise estimation with the well-trained network It can effectively suppress noise while preserving phase fringe edges.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
