Resolving Interferometric SAR Speckle Problem Using Adaptive GAN

Abstract

The interference of reflected waves from multiple elementary scatterers produces speckle, which appears as a granular noise in synthetic aperture radar (SAR) images. These speckles in SAR images cause difficulty in image interpretation, which reduces the effectiveness of image segmentation and classification. In this paper, we propose an effective solution using generative adversarial networks (GAN) to decrease speckle noise while preserving texture features.The convolutional block attention module (CBAM) boosts the essential features in upsampled data and supports the creation of a final denoised real InSAR data.The proposed deep-learning method proved that CBAM enhanced the peak signal-to-noise ratio using the GAN. Furthermore, CBAM improved structural similarity (SSIM) to 99 and achieved the minimum mean squared error. The despeckle performance was enhanced using GAN-ResUNet in which the SSIM equals 0.999. The denoising performance proved that the use of GAN with CBAM as the generator inspired by ResUNeT achieved the best performance compared to other experiments results.

Convolutional deep learning network for InSAR phase denoising and unwrapping

“…Path-following algorithms are difficult to deal with the integration-path inconsistency issue. Recently, deep learning methods have been widely exploited in PhU [16][17][18][19][20]. Such methods include encoder-decoder [16][17] architectures to achieve semantic segmentation.…”

“…Such methods include encoder-decoder [16][17] architectures to achieve semantic segmentation. Other approaches combine additional factors such as the gradient information [18], the least squares [19] and the residual learning [20] to the original deep learning methods to improve the final unwrapped phase reconstruction. In the field of InSAR data, the unwrapping problem becomes more challenging due to two main reasons: the complex wrapped phase features and the high noise factor.…”