Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PolMERLIN: Self-Supervised Polarimetric Complex SAR Image Despeckling with Masked Networks

Shunya Kato, Masaki Saito, Katsuhiko Ishiguro, Sol Cummings

TL;DR

PolMERLIN tackles the challenge of despeckling multi-polarization complex SAR imagery without ground-truth by extending a self-supervised framework to utilize cross-polarization correlations. It introduces channel masking and spatial masking to form a Noise2Noise-based objective, and implements them within a U-Net despeckling network to recover masked real/imaginary components across HH and VV channels. The method is validated on synthetic gamma-speckle data and real TerraSAR-X dual-polarization imagery, showing superior PSNR/SSIM and ENL gains over MERLIN and supervised baselines, with spatial masking providing additional improvements. This work enhances despeckling performance for modern multi-polarization SAR systems, facilitating better downstream analysis while avoiding reliance on ground-truth clean images.

Abstract

Despeckling is a crucial noise reduction task in improving the quality of synthetic aperture radar (SAR) images. Directly obtaining noise-free SAR images is a challenging task that has hindered the development of accurate despeckling algorithms. The advent of deep learning has facilitated the study of denoising models that learn from only noisy SAR images. However, existing methods deal solely with single-polarization images and cannot handle the multi-polarization images captured by modern satellites. In this work, we present an extension of the existing model for generating single-polarization SAR images to handle multi-polarization SAR images. Specifically, we propose a novel self-supervised despeckling approach called channel masking, which exploits the relationship between polarizations. Additionally, we utilize a spatial masking method that addresses pixel-to-pixel correlations to further enhance the performance of our approach. By effectively incorporating multiple polarization information, our method surpasses current state-of-the-art methods in quantitative evaluation in both synthetic and real-world scenarios.

PolMERLIN: Self-Supervised Polarimetric Complex SAR Image Despeckling with Masked Networks

TL;DR

PolMERLIN tackles the challenge of despeckling multi-polarization complex SAR imagery without ground-truth by extending a self-supervised framework to utilize cross-polarization correlations. It introduces channel masking and spatial masking to form a Noise2Noise-based objective, and implements them within a U-Net despeckling network to recover masked real/imaginary components across HH and VV channels. The method is validated on synthetic gamma-speckle data and real TerraSAR-X dual-polarization imagery, showing superior PSNR/SSIM and ENL gains over MERLIN and supervised baselines, with spatial masking providing additional improvements. This work enhances despeckling performance for modern multi-polarization SAR systems, facilitating better downstream analysis while avoiding reliance on ground-truth clean images.

Abstract

Despeckling is a crucial noise reduction task in improving the quality of synthetic aperture radar (SAR) images. Directly obtaining noise-free SAR images is a challenging task that has hindered the development of accurate despeckling algorithms. The advent of deep learning has facilitated the study of denoising models that learn from only noisy SAR images. However, existing methods deal solely with single-polarization images and cannot handle the multi-polarization images captured by modern satellites. In this work, we present an extension of the existing model for generating single-polarization SAR images to handle multi-polarization SAR images. Specifically, we propose a novel self-supervised despeckling approach called channel masking, which exploits the relationship between polarizations. Additionally, we utilize a spatial masking method that addresses pixel-to-pixel correlations to further enhance the performance of our approach. By effectively incorporating multiple polarization information, our method surpasses current state-of-the-art methods in quantitative evaluation in both synthetic and real-world scenarios.
Paper Structure (14 sections, 5 equations, 4 figures, 2 tables)

This paper contains 14 sections, 5 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Statistical Model of Complex SAR
  3. Proposed Method
  4. Statistical Model of Polarimetric Complex SAR
  5. Channel Masking
  6. PolMERLIN
  7. Experiments
  8. Procedure
  9. Synthetic Speckle Noise
  10. Real SAR Images
  11. Training Details
  12. Experiments on Synthetic Speckle Noise
  13. Experiments on Real SAR
  14. Conclusion

Figures (4)

  • Figure 1: Comparison of self-supervised despeckling methods. Blue pixels indicate the range of receptive fields. (a) Speckle2Void uses blind-spotting (spatial information). (b) MERLIN predicts a Re/Im component from the counterpart component in a single polarization signal. (c) The proposed PolMERLIN predicts Re/Im components from the counterpart components of multi-polarization signals. To use spatial information as well, PolMERLIN intentionally removes some pixels from the receptive field.
  • Figure 2: Overview of PolMERLIN. After channel masking and spatial masking are performed on the input, the model is trained by restoring the masked components using a despeckling network. For the despeckling network architecture we used the U-Net model Wang2022.
  • Figure 3: Qualitative results for the R channel of the BSDS500 image despeckled with pseudo-noise.
  • Figure 9: Qualitative despeckling results of MERLIN and Ours(c+s) in TerraSAR-X. Noisy bright (white) spots are found in the lower-left part of the MERLIN result. Ours(c+s)' result is not suffered from such artifacts.