[Paper Review] SAR Image Colorization: Converting Single-Polarization to Fully Polarimetric Using Deep Neural Networks

Qian Song, Feng Xu+1 more|arXiv (Cornell University)|Jul 22, 2017

Synthetic Aperture Radar (SAR) Applications and Techniques1 references3 citations

TL;DR

This paper proposes a deep learning framework that converts single-polarization SAR images into fully polarimetric SAR images using a two-stage network: a feature extractor captures multi-scale spatial features, followed by a feature translator that maps them to polarimetric features for covariance matrix reconstruction. The method achieves high fidelity reconstruction, enabling direct application of PolSAR analysis techniques like decomposition and classification.

ABSTRACT

A deep neural networks based method is proposed to convert single polarization grayscale SAR image to fully polarimetric. It consists of two components: a feature extractor network to extract hierarchical multi-scale spatial features of grayscale 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 fully polarimetric 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. Existing PolSAR applications such as model-based decomposition and unsupervised classification can be applied directly to the reconstructed full-pol SAR images. This framework can be easily extended to reconstruction of full-pol data from compact-pol data. The experiment results also show that the proposed method could be potentially used for interference removal on the cross-polarization channel.

Motivation & Objective

• To address the scarcity of fully polarimetric SAR (PolSAR) data by synthesizing it from single-polarization SAR images.
• To develop a deep neural network framework that learns to predict full polarimetric covariance matrices from intensity-only SAR images.
• To enable downstream PolSAR applications such as model-based decomposition and unsupervised classification on reconstructed data.
• To explore potential for interference removal in cross-polarization channels through the learned mapping.
• To extend the method to reconstruction from compact-pol SAR data, broadening its applicability.

Proposed method

• A two-stage deep neural network is employed: first, a feature extractor network processes the input single-pol SAR image to extract hierarchical multi-scale spatial features.
• The feature extractor uses convolutional layers to learn spatial representations from the grayscale SAR intensity data.
• A subsequent feature translator network maps the extracted spatial features to polarimetric features corresponding to the full covariance matrix of each pixel.
• The network is trained end-to-end to minimize reconstruction error between predicted and ground-truth covariance matrices.
• The predicted covariance matrices are used to reconstruct a fully polarimetric SAR image, effectively 'colorizing' the single-pol input.
• The framework is adaptable to compact-pol SAR data by modifying the input and output layers to match the compact-pol signal model.

Experimental results

Research questions

• RQ1 Can a deep neural network accurately reconstruct the full polarimetric covariance matrix from a single-polarization SAR intensity image?
• RQ2 To what extent does the reconstructed PolSAR data preserve the statistical and structural properties of real full-pol SAR data?
• RQ3 Can standard PolSAR analysis techniques, such as model-based decomposition and unsupervised classification, be effectively applied to the reconstructed data?
• RQ4 Does the method show potential for mitigating cross-polarization channel interference in SAR data?
• RQ5 Can the framework be generalized to reconstruct full-pol data from compact-pol SAR inputs?

Key findings

• The proposed method successfully reconstructs fully polarimetric SAR images from single-pol inputs with high fidelity, as validated by qualitative and quantitative comparisons with real PolSAR data.
• The reconstructed covariance matrices closely match the ground truth, enabling accurate application of standard PolSAR analysis tools such as model-based decomposition.
• Unsupervised classification applied to the reconstructed data achieves performance comparable to that on real full-pol SAR data.
• The method demonstrates robustness in preserving polarimetric diversity and scattering characteristics across different terrain types.
• The framework shows potential for interference suppression in cross-polarization channels by learning clean polarimetric representations from noisy or corrupted inputs.
• The approach is extensible to compact-pol SAR data, suggesting broader applicability in scenarios with limited polarimetric acquisition.