Radar Image Colorization: Converting Single-Polarization to Fully Polarimetric Using Deep Neural Networks

Numerous radar polarimetry theories and polarimetric synthetic aperture radar (PolSAR) processing methods have been developed. However, the vast majority of SAR images are not fully polarimetric (full-pol). This paper proposes "radar image colorization" to reconstruct a full-pol image from a non-full-pol image, so that existing PolSAR methods, such as model-based decomposition and unsupervised classification, can be directly applied to the reconstructed full-pol SAR images. It proposes to train a specially designed deep neural network to convert a single polarization gray-scale SAR image to full-pol. It consists of two components: a feature extractor network to extract hierarchical multi-scale spatial features of the gray-scale SAR image, followed by a feature translator network to map spatial feature to polarimetric feature with which the polarimetric covariance matrix of each pixel can be reconstructed. Both qualitative and quantitative experiments with real full-pol data are conducted to show the efficacy of the proposed method. The reconstructed full-pol SAR image agrees well with the true full-pol image, not only in the sense of visual similarity but also in the sense of real PolSAR applications, such as target decomposition and terrain classification.