CBIPDNet: A Novel Method for InSAR Deformation Interferometric Phase Filtering Using Deep Learning Network

Abstract

The denoising of phase is a crucial process that impacts the accuracy of data processing in differential interferometric synthetic aperture radar. Especially in the area of large-gradient deformation, the phase filtering method is very easy to cause phase losses. This has a significant impact on the final deformation acquisition. To address this issue, here, a deep convolutional blind denoising network-based interferometric phase filtering method, named CBIPDNet, is proposed. Different from the previously proposed deep learning phase filtering methods, CBIPDNet does not add noise to the input before filtering, but adds noise to the input during the training process. Furthermore, CBIPDNet uses a CNN structure for adaptive noise estimation and uses a residual module for nonblind filtering. Therefore, CBIPDNet can be considered as an adaptive phase filtering algorithm. More importantly, the added noise is composed of heteroscedastic Gaussian noise + simulated real noise of the imaging process, which is closer to the real interferometric noise phase. Moreover, the denoising effect of targets of different scales through the asymmetric loss function has been significantly improved, which can improve the detail preservation ability of regions with substantial gradient deformations. The experimental results demonstrate that CBIPDNet is capable of enhancing phase quality and increasing phase unwrapping accuracy compared to the current interferometric filtering methods.

Keywords

• Convolutional blind denoising network
• differential interferometric synthetic aperture radar (DInSAR)
• phase filtering
• subsidence deformation