PolSAR Band-to-Band Image Translation Using Conditional Adversarial Networks

Abstract

PolSAR image captured at different frequency bands contains varied information of the same target object. It has been reported that multi-frequency PolSAR data incurs high acquisition costs and computational requirements. In this paper, we put forward a novel concept of PolSAR band-to-band image translation to synthesize multi-frequency PolSAR images from a single frequency PolSAR image. Our proposed method uses PolSAR images captured at a particular frequency to generate its representation in different frequency bands based on image representation and target understanding. We leverage a deep neural network, particularly conditional adversarial network to perform the task. Our proposed framework shows promising results on AIRSAR dataset both qualitatively in terms of visual similarity and quantitatively in terms of root mean square error(RMSE).