Multi-Frequency Polarimetric SAR Classification Based on Riemannian Manifold and Simultaneous Sparse Representation

Fan Yang, Wei Gao, Bin Xu, Jian YangRemote 2015 · DOI

TL;DR

This paper introduces a novel approach for classifying land cover using polarimetric Synthetic Aperture Radar (PolSAR) data. By leveraging the geometric properties of PolSAR data (which reside on a Riemannian manifold) and combining it with sparse representation techniques using a specialized "Stein kernel," the authors achieve more accurate land cover classification. They further extend this to multi-frequency data, demonstrating significant improvements over traditional methods.

What Problem Does This Paper Address?

Classifying land cover from polarimetric SAR (PolSAR) data is a challenging task. PolSAR data, represented by covariance or coherency matrices, inherently lives on a Riemannian manifold, not a standard Euclidean space. Traditional classification methods, especially those relying on Euclidean metrics like standard sparse representation, struggle with this geometric structure. This leads to suboptimal performance, particularly in complex land cover scenarios. Furthermore, combining information from different SAR frequency bands (multi-frequency PolSAR) is known to improve classification accuracy, but effectively merging this information while respecting the underlying data structure has been an open problem.

Methods

The core innovation lies in two main areas:

  • Stein-SRC on Riemannian Manifolds: Instead of using standard Euclidean distances, the authors employ a "Stein kernel" to measure similarity between PolSAR data points (represented as matrices) on their native Riemannian manifold. This Stein kernel is combined with sparse representation-based classification (SRC). This approach, termed Stein-SRC, directly handles the geometric properties of the data without requiring complex transformations or assumptions about specific distributions.
  • Simultaneous Stein-SRC for Multi-Frequency Data: To leverage multi-frequency PolSAR data, the authors extend Stein-SRC using simultaneous sparse representation. This method assumes that the sparse representations across different frequency bands are correlated. By jointly optimizing the sparse representations for multiple bands, they can effectively merge information and achieve better classification, especially when different bands have complementary strengths.

Datasets

The proposed methods were evaluated on two primary datasets:

  • San Francisco, USA: Data from the JPL AIRSAR sensor was used for single-band and multi-band classification experiments. This dataset features four land cover types: sea, urban areas, forests, and bare ground.
  • Flevoland, Netherlands: Data from the EMISAR sensor (P, L, and C bands) was used for multi-frequency classification. This dataset contains 14 types of crops, providing a more complex classification scenario.

Key Insights and Ideas

  • Riemannian Geometry is Crucial: The paper strongly emphasizes that treating PolSAR data as points in a Euclidean space is fundamentally flawed. Utilizing Riemannian geometry and appropriate metrics (like the Stein kernel) is essential for accurate classification.
  • Sparse Representation Benefits from Geometric Awareness: Standard sparse representation, while powerful, needs to be adapted to non-Euclidean data. The Stein kernel provides a principled way to do this for PolSAR data.
  • Multi-Frequency Data Synergy: Combining information from different SAR frequencies significantly boosts classification accuracy. The simultaneous sparse representation approach effectively captures this synergy by assuming correlated sparse representations across bands.
  • Efficiency vs. Accuracy Trade-off: The authors also propose "simplified" versions of their methods (simplified Stein-SRC and simplified simultaneous Stein-SRC) that reduce computational complexity at a slight cost to accuracy, making them suitable for real-time applications.

Key Results

  • Improved Accuracy: Stein-SRC consistently outperformed the traditional Wishart classifier on both single-band (San Francisco) and multi-band (Flevoland) datasets. For the San Francisco data, Stein-SRC achieved an overall accuracy of 93.3% compared to the Wishart classifier's 87.0%.
  • Multi-Frequency Advantage: Simultaneous Stein-SRC demonstrated superior performance when combining P, L, and C bands, achieving an overall accuracy of 99.0% for the Flevoland dataset, significantly outperforming the Wishart classifier (96.1%).
  • Robustness: Stein-SRC showed less performance degradation with insufficient training samples compared to some distribution-based models.
  • Computational Efficiency: While Stein-SRC is computationally intensive (especially the Lasso algorithm), simplified versions offer a good balance between accuracy and speed. Simplified Stein-SRC was faster than the traditional Wishart classifier in some cases.

Limitations and Future Work

  • Computational Cost: The full Stein-SRC and simultaneous Stein-SRC methods, particularly due to the Lasso algorithm, can be computationally demanding, especially for large datasets or real-time applications. The authors acknowledge this and propose simplified versions.
  • Parameter Selection: While the paper discusses the influence of the kernel parameter $\sigma$, further research could explore more automated or adaptive methods for parameter selection.
  • Generalizability: While tested on two diverse datasets, further validation on a wider range of land cover types and sensor data would be beneficial.
  • Future Directions: The authors suggest that the proposed framework could be extended to other applications involving data on Riemannian manifolds, such as image segmentation or change detection. Further exploration of efficient algorithms for sparse representation on manifolds is also a promising avenue.

Summary updated

Abstract

Normally, polarimetric SAR classification is a high-dimensional nonlinear mapping problem. In the realm of pattern recognition, sparse representation is a very efficacious and powerful approach. As classical descriptors of polarimetric SAR, covariance and coherency matrices are Hermitian semidefinite and form a Riemannian manifold. Conventional Euclidean metrics are not suitable for a Riemannian manifold, and hence, normal sparse representation classification cannot be applied to polarimetric SAR directly. This paper proposes a new land cover classification approach for polarimetric SAR. There are two principal novelties in this paper. First, a Stein kernel on a Riemannian manifold instead of Euclidean metrics, combined with sparse representation, is employed for polarimetric SAR land cover classification. This approach is named Stein-sparse representation-based classification (SRC). Second, using simultaneous sparse representation and reasonable assumptions of the correlation of representation among different frequency bands, Stein-SRC is generalized to simultaneous Stein-SRC for multi-frequency polarimetric SAR classification. These classifiers are assessed using polarimetric SAR images from the Airborne Synthetic Aperture Radar (AIRSAR) sensor of the Jet Propulsion Laboratory (JPL) and the Electromagnetics Institute Synthetic Aperture Radar (EMISAR) sensor of the Technical University of Denmark (DTU). Experiments on single-band and multi-band data both show that these approaches acquire more accurate classification results in comparison to many conventional and advanced classifiers.
BibTeX
@inproceedings{yang2015multi,
  author = {Fan Yang and Wei Gao and Bin Xu and Jian Yang},
  title = {{Multi-Frequency Polarimetric SAR Classification Based on Riemannian Manifold and Simultaneous Sparse Representation}},
  booktitle = {Remote},
  year = {2015},
  doi = {10.3390/rs70708469},
}
Figure