Abstract
Target detection in Polarimetric Synthetic Aperture Radar (PolSAR) imagery is often complicated by speckle noise and the presence of complex natural backgrounds. This paper introduces a novel detection framework that exploits the physical property of reflection symmetry to distinguish man-made objects from natural clutter. The method integrates SIFT-based feature fusion of specific polarization terms ($C_{12}$ and $C_{23}$) with an iterative patch-level Wishart classifier. Experimental results demonstrate significant improvements in detection accuracy and robustness, particularly in low target-to-clutter ratio (TCR) environments.
Introduction
The detection of man-made targets, such as ships and vehicles, in SAR images is a critical task for maritime and terrestrial surveillance. However, traditional single-channel SAR methods often struggle with "speckle noise" and the difficulty of separating targets from complex natural backgrounds like forests or sea surfaces.
The core innovation of this research lies in the exploitation of reflection symmetry. Natural surfaces generally possess reflection symmetry, which causes the off-diagonal terms $C_{12}$ and $C_{23}$ in the polarimetric covariance matrix to tend toward zero. Conversely, man-made structures, characterized by complex dihedral and trihedral shapes, break this symmetry. The proposed framework focuses on effectively combining these symmetry-breaking terms to generate high-contrast imagery suitable for automated detection.
Methodology
The proposed approach consists of two primary stages: SIFT-based image fusion and an iterative Wishart classification.
SIFT-Based Image Fusion
The polarization terms $r_1 = |C_{12}|$ and $r_2 = |C_{23}|$ contain complementary structural information. To fuse these terms effectively, the authors employ the Scale-Invariant Feature Transform (SIFT) to detect "blob-like" structures representing potential targets. An "index map" is generated from these SIFT key points. In regions identified as potential targets, a MAX operation is applied to preserve the strongest signal. In background regions, a MIN operation is used to suppress clutter, resulting in a cleaner fused image.
Iterative Wishart Classifier
The fused image is processed using an unsupervised classifier based on the Wishart distance, a statistical metric optimized for PolSAR data distributions. Unlike standard pixel-level classifiers, this method operates at the patch level (using $3 \times 3$ windows), enhancing robustness to noise. The classifier iteratively reassigns pixels to "target" or "clutter" classes until convergence is achieved.
Experimental Results
The framework was validated using the AFRL Airborne X-band SAR dataset and the RADARSAT-2 Spaceborne C-band SAR dataset.
Key performance metrics include:
- High Accuracy: On the AFRL dataset, the fusion method achieved an Area Under the Curve (AUC) of 0.9507, significantly outperforming standard CFAR detectors (0.8587).
- Robustness: Synthetic data testing showed the method maintains high performance even at a Target-to-Clutter Ratio (TCR) as low as 2dB, where traditional methods like SPAN or PWF typically fail.
- Visual Clarity: Qualitative analysis confirms that the fused images effectively suppress background clutter while maintaining bright, well-defined target contours.
Discussion and Future Work
By utilizing reflection symmetry as a physics-based filter, the researchers have developed a robust method for highlighting complex man-made structures. The integration of SIFT for denoising and the patch-level Wishart classifier provides a significant advantage in low-contrast scenarios.
However, limitations remain. False alarms can still occur in coastal areas with complex scattering mechanisms, and very small targets remain difficult to distinguish from noise. Future research will explore the inclusion of attribute scattering center information to better characterize the geometry of wave reflections, further improving the distinction between small targets and random clutter.
Summary updated