Exploiting GAN-Based SAR to Optical Image Transcoding for Improved Classification via Deep Learning

Learning the proxy task of transcoding SAR images into optical images forces an employed conditional generative adversarial network (GAN) to distinguish between different land surfaces. Such a network can then be used to build a classifier with significantly fewer free parameters that generalizes well even when trained on a very small amount of labeled data. We train such a GAN on aligned Sentinel-1 and Sentinel-2 image pairs. We then show that a pre-trained classifier using these features learned from transcoding outperforms classifiers that are trained from scratch when only a very limited amount of labeled pixels are available for training.