Contribution of Information Geometry for Polarimetric SAR Classification in Heterogeneous Areas

Jean‐Philippe Ovarlez, Pierre Formont, Frédéric Pascal, Gabriel Vasile, Laurent Ferro-Famil2014

TL;DR

This paper introduces a novel approach to classifying Polarimetric Synthetic Aperture Radar (PolSAR) images by employing information geometry. By using a Riemannian mean instead of a traditional Euclidean mean to define class centers, the authors achieve more stable and interpretable classifications, particularly in complex, heterogeneous areas.

What Problem Does This Paper Address?

PolSAR systems generate high-resolution images that capture detailed scattering information, enabling the discrimination of different surface features. A common method for classifying these images is the H/α (entropy-alpha) classification, which relies on analyzing the eigenvalues of the polarimetric covariance matrix (CM). However, traditional classification algorithms often use the Sample Covariance Matrix (SCM) and a simple Euclidean mean to calculate class centers. This approach can lead to instability and difficulty in interpreting the results, especially in heterogeneous areas where scattering mechanisms are mixed. The paper addresses the challenge of accurately and stably classifying these complex scenes by proposing a more robust method for defining class centers.

Methods

The core innovation of this paper lies in the application of information geometry to the classification process. Instead of using the standard Euclidean (arithmetic) mean to compute the average CM for each class, the authors propose using the Riemannian geometric mean. This is a more appropriate metric for the space of covariance matrices, which is not Euclidean.

The classification process itself is based on an iterative algorithm, similar to K-means clustering. Each pixel is represented by its CM. The algorithm iteratively:

  1. Assigns pixels to classes: This is done using a distance metric between the pixel's CM and the class centers. The paper discusses both the Wishart distance for Gaussian assumptions and a derived distance for the Spherically Invariant Random Vector (SIRV) model.
  2. Recomputes class centers: This is where the key difference lies. Traditionally, an arithmetic mean of the CMs within a class is used. Here, the Riemannian geometric mean is computed. This involves minimizing a specific objective function (Equation 6 in the paper) that leverages the geometry of the space of positive-definite matrices.

The paper also mentions the use of the Fixed Point Estimate (FPE) for computing the covariance matrix, which is presented as an improvement over the standard SCM, especially for textured areas.

Datasets

The paper evaluates its method on two real-world PolSAR datasets:

  • A dataset acquired in Brétigny, France by the ONERA RAMSES system (X-band).
  • A dataset acquired in Paracou, French Guyana by the ONERA SETHI system (UHF-band).

Both datasets have a ground resolution of 1.3x1.3m. No new benchmark datasets are introduced.

Key Insights and Ideas

  • The space of covariance matrices is not Euclidean: This fundamental insight drives the need for a more sophisticated mean calculation than the simple arithmetic average.
  • Riemannian geometric mean leads to more stable class centers: By accounting for the intrinsic geometry of the CM space, the class centers are less prone to drifting significantly during iterations, leading to more consistent and interpretable results.
  • Improved classification in heterogeneous areas: The combination of the FPE for CM estimation and the Riemannian mean for class centers yields better results in complex, highly textured scenes where traditional methods struggle.
  • Potential for refined interpretation: The stability of class centers allows for a more meaningful interpretation of the H/α/A classification, as the centers are less likely to be artifacts of the averaging process.

Key Results

The paper presents qualitative results comparing classifications using the arithmetic mean versus the geometric mean.

  • Figure 2 (Brétigny dataset) shows that using the geometrical mean makes corner reflectors (class 8) stand out more distinctly. Urban areas (classes 1, 3, 6, 7) are also better represented, while fields and forests (remaining classes) are more homogeneous.
  • Figure 3 (Paracou dataset) demonstrates that the geometrical mean refines the classification of specific areas. Notably, the number of pixels in class 8 (which represents a road and an unidentified area) is reduced compared to the arithmetic mean. Pixels that left class 8 were re-attributed to neighboring classes (4 and 5).
  • The paper notes that using the Riemannian distance (Eq. 7) in conjunction with the geometric mean further refines the classification.

While quantitative metrics like accuracy percentages are not explicitly provided, the visual results suggest a clear improvement in classification detail and stability.

Limitations and Future Work

  • Ground Truth: The authors acknowledge that detailed ground truth for the Paracou dataset is limited, making a definitive interpretation of some classification results challenging.
  • Quantification: The paper states that "Simulations will be done in order to quantify the contribution of this method." This suggests that a more thorough quantitative analysis is planned for future work.
  • Further Exploration: The authors express that "using geometry of information brings additional information" and that "First results using the Riemannian metric for PolSAR classification are encouraging and suggest an improvement of the classifications by using these methods more intently." This indicates a strong motivation for further research and application of these geometric approaches.

Summary updated

Abstract

We discuss in the paper the use of the Riemannian mean given by the differential geometric tools. This geometric mean is used in this paper for computing the centers of class in the polarimetric H/α unsupervised classification process. We can show that the centers of class will remain more stable during the iteration process, leading to a different interpretation of the H/α/A classification. This technique can be applied both on classical SCM and on Fixed Point covariance matrices. Used jointly with the Fixed Point CM estimate, this technique can give nice results when dealing with high resolution and highly textured polarimetric SAR images classification. 1.
BibTeX
@misc{ovarlez2014contribution,
  author = {Jean‐Philippe Ovarlez and Pierre Formont and Frédéric Pascal and Gabriel Vasile and Laurent Ferro-Famil},
  title = {{Contribution of Information Geometry for Polarimetric SAR Classification in Heterogeneous Areas}},
  year = {2014},
}
Figure