A new clustering algorithm for PolSAR images segmentation

Violeta Roizman, Gordana Drašković, Frédéric Pascal2019 · DOI

TL;DR

This paper introduces a novel robust clustering algorithm, R-EM, for segmenting PolSAR (Polarimetric Synthetic Aperture Radar) images, which are notoriously noisy and non-Gaussian. By incorporating a flexible texture parameter and robust estimators, R-EM significantly outperforms traditional methods like k-means and Gaussian Mixture Models (GMM) in simulated scenarios, offering more accurate and less noisy segmentation.

What Problem Does This Paper Address?

Segmenting PolSAR images is crucial for various applications, such as land use classification. However, PolSAR data is inherently complex due to speckle noise and often exhibits non-Gaussian characteristics, making traditional clustering algorithms struggle. Standard methods like k-means and Gaussian Mixture Models (GMM) assume data follows a Gaussian distribution and are sensitive to outliers and noise, leading to inaccurate segmentations. This paper aims to address this by developing a more robust clustering algorithm that can effectively handle the non-Gaussian nature and noise present in PolSAR imagery.

Methods

The proposed algorithm, R-EM (Robust Expectation-Maximization), builds upon the classical EM algorithm but introduces two key innovations:

  • Robust Estimators: Instead of relying on standard sample means and covariance matrices, R-EM uses Tyler's M-estimators. These are known to be more robust to outliers and deviations from the assumed distribution, which is critical for noisy PolSAR data.
  • Flexible Texture Parameter: The algorithm incorporates a texture parameter ($\tau_{ik}$) for each pixel and cluster. This parameter allows for greater flexibility in modeling the data distribution, moving beyond the strict Gaussian assumption. The estimation of this parameter is done using a trimmed mean over local patches, which helps to discard extreme values and further enhances robustness.

The algorithm iteratively estimates cluster parameters (mean, covariance, and proportions) and the texture parameter for each pixel, aiming to maximize a likelihood function that accounts for these robust estimators and the flexible texture model.

Datasets

The paper evaluates the proposed algorithm on two types of datasets:

  • Simulated PolSAR Data: Generated using a Markov Random Field (MRF) model with assigned polarimetric signatures and Gaussian speckle noise. These simulated datasets are created to mimic real PolSAR data characteristics and are generated at different "look" levels (6, 9, and 12-looked), representing varying degrees of noise reduction.
  • Real PolSAR Data: A fragment of a real PolSAR image is used for evaluation. This fragment is slightly smoothed before segmentation.

Key Insights and Ideas

  • The Importance of Robustness: The paper strongly emphasizes that the non-Gaussian and noisy nature of PolSAR data necessitates robust statistical methods. Standard EM algorithms, while powerful, are brittle when these assumptions are violated.
  • Texture as a Key to Flexibility: Introducing a texture parameter that varies per pixel and cluster is a significant step towards modeling the complex variability in PolSAR images. This allows the algorithm to adapt to local data characteristics more effectively.
  • Patch-Based Processing for Local Structure: Using local patches for texture estimation leverages the spatial context of the image, which is crucial for accurate segmentation and helps in smoothing noisy amplitude values.
  • Trimmed Mean for Outlier Rejection: The use of a trimmed mean for texture estimation is a clever way to mitigate the impact of remaining extreme values even after initial smoothing.

Key Results

  • Simulated Data Performance: R-EM consistently outperforms k-means and GMM on simulated PolSAR data across different look numbers. For instance, clustering accuracy for 12-looked data reaches 0.99 with R-EM, compared to 0.96 for k-means and 0.98 for GMM. The segmented images produced by R-EM are visibly less noisy and more coherent.
  • Real Data Performance: On real PolSAR data, the results are more nuanced. While R-EM and GMM both perform better than k-means, direct comparison between R-EM and GMM is less clear-cut, with each method showing advantages in different regions of the image. This suggests that while R-EM is generally more robust, the specific characteristics of real data can lead to more complex outcomes.

Limitations and Future Work

  • Limitations: The authors acknowledge that on real PolSAR data, the results were "mitigated" compared to simulations, indicating that further tuning of parameters is necessary. The current work focuses on amplitude real values, while PolSAR data is inherently complex-valued.
  • Future Work: The authors suggest several promising directions:
  • Exploring the Complex Case: Extending the algorithm to directly handle complex-valued PolSAR data by deriving new estimators from the complex likelihood.
  • Deterministic $\tau$ Estimation: Developing an algorithm that does not assume a deterministic $\tau$ parameter, further enhancing flexibility.
  • Parameter Tuning: Investigating strategies for optimal parameter tuning, especially for real-world applications.

Summary updated

Abstract

International audience
BibTeX
@misc{roizman2019new,
  author = {Violeta Roizman and Gordana Drašković and Frédéric Pascal},
  title = {{A new clustering algorithm for PolSAR images segmentation}},
  year = {2019},
  doi = {10.1109/camsap45676.2019.9022495},
}
Figure