[Paper Review] Despeckling Polarimetric SAR Data Using a Multi-Stream Complex-Valued Fully Convolutional Network

Adugna Mullissa, Claudio Persello+1 more|arXiv (Cornell University)|Mar 12, 2021

Synthetic Aperture Radar (SAR) Applications and Techniques22 references29 citations

TL;DR

This paper proposes CV-deSpeckNet, a multi-stream complex-valued fully convolutional network that directly processes the complex-valued covariance matrix of dual-polarimetric SAR data, modeling speckle as additive noise in the log domain. By learning joint real-imaginary representations without separating components, it achieves superior despeckling performance with fewer training samples and higher generalization than real-valued counterparts and state-of-the-art methods, yielding higher PSNR, SSIM, and ENL values while preserving polarimetric features.

ABSTRACT

A Polarimetric Synthetic Aperture Radar (PolSAR) sensor is able to collect images in different polarization states, making it a rich source of information for target characterization. PolSAR images are inherently affected by speckle. Therefore, before deriving ad hoc products from the data, the polarimetric covariance matrix needs to be estimated by reducing speckle. In recent years, deep learning based despeckling methods have started to evolve from single channel SAR images to PolSAR images. To this aim, deep learning based approaches separate the real and imaginary components of the complex-valued covariance matrix and use them as independent channels in a standard convolutional neural networks. However, this approach neglects the mathematical relationship that exists between the real and imaginary components, resulting in sub-optimal output. Here, we propose a multi-stream complex-valued fully convolutional network to reduce speckle and effectively estimate the PolSAR covariance matrix. To evaluate the performance of CV-deSpeckNet, we used Sentinel-1 dual polarimetric SAR images to compare against its real-valued counterpart, that separates the real and imaginary parts of the complex covariance matrix. CV-deSpeckNet was also compared against the state of the art PolSAR despeckling methods. The results show CV-deSpeckNet was able to be trained with a fewer number of samples, has a higher generalization capability and resulted in a higher accuracy than its real-valued counterpart and state-of-the-art PolSAR despeckling methods. These results showcase the potential of complex-valued deep learning for PolSAR despeckling.

Motivation & Objective

• To address the limitation of real-valued deep learning methods that separate real and imaginary parts of the complex covariance matrix, neglecting their mathematical relationship.
• To develop a complex-valued neural network architecture that preserves the intrinsic structure of PolSAR covariance matrices during despeckling.
• To improve generalization and reduce training data requirements by leveraging complex-valued feature learning in a multi-stream fully convolutional framework.
• To outperform existing state-of-the-art PolSAR despeckling methods in both quantitative metrics and visual quality.

Proposed method

• The method models the log-transformed covariance matrix as an additive noise model, enabling the use of fully convolutional networks (FCNs) to estimate clean covariance and noise components separately.
• A multi-stream architecture uses two identical complex-valued FCNs: one for estimating the clean covariance matrix (FCNcov), and another for estimating the noise (FCNnoise).
• The final estimate is reconstructed by exponentiating the sum of the clean covariance and noise predictions, restoring the original multiplicative noise model.
• The network is trained end-to-end using paired noisy and reference (temporally averaged) covariance matrices, with loss functions optimized for PSNR and SSIM.
• The complex-valued operations preserve phase and amplitude relationships between real and imaginary components, enabling more robust feature learning.
• During inference, only the clean covariance estimation branch (FCNcov) is used, ensuring efficient deployment.

Experimental results

Research questions

• RQ1 Can a complex-valued deep learning architecture outperform real-valued networks in despeckling dual-polarimetric SAR covariance matrices?
• RQ2 Does modeling the covariance matrix in the complex domain, without separating real and imaginary parts, lead to better generalization and lower data requirements?
• RQ3 How does the proposed multi-stream complex-valued FCN compare to state-of-the-art PolSAR despeckling methods in terms of quantitative metrics and visual quality?
• RQ4 To what extent can the model generalize to unseen test scenes without fine-tuning, given limited training data?

Key findings

• CV-deSpeckNet achieved the highest PSNR (31.12) and SSIM (0.87) on test image 1, outperforming all baselines including the real-valued RV-FCN and state-of-the-art methods.
• On test image 2, CV-deSpeckNet achieved a PSNR of 37.28 and SSIM of 0.88 when fine-tuned on the reference image, approaching the upper bound of performance.
• The model achieved an ENL of 177.58 on test image 1 and 86.81 on test image 2, significantly outperforming the next best method (deSpeckNet: 81.46 and 29.08).
• CV-deSpeckNet reduced the absolute error in polarimetric parameters: α (˜α) by 57% on test image 1 and 70% on test image 2 compared to the next best method.
• The model generalized effectively to new scenes without fine-tuning, achieving 90% of the performance of a fully tuned network on test image 1 and 92% on test image 2.
• CV-deSpeckNet required only 7.01 hours to train, compared to 1.3 hours for RV-FCN and 2.6 hours for deSpeckNet, indicating efficient learning despite complex-valued operations.