A Novel Approach for SAR to Optical Image Registration using Deep Learning

Image registration is one of the essential pre-processing steps in remote sensing data applications. Since Synthetic Aperture Radar (SAR) and optical images give complementary information, they are used simultaneously for several applications. However, they differ in their geometric and radiometric properties; hence, SAR to optical image registration is challenging. An image registration approach based on Conditional Generative Adversarial Network (cGAN) to synthesize optical-like SAR images and Fast Fourier Transform (FFT) based correlation is proposed in this work. cGAN does the image translation of the SAR patch and generates a corresponding optical-like patch. The optical-like SAR image is used as a reference to register the optical image. Sentinel-l SAR and Sentine1-2 optical patch pairs are used for this work. When registered with optical-like patches using FFTbased correlation, optical patches could give a sub-pixel accuracy.