Abstract
A novel despeckling technique for Synthetic Aperture Radar (SAR) images is presented, leveraging multiple image reconstruction and selective 3-dimensional filtering. By applying Compressive Sensing (CS) theory to multiple subsets of pixels derived from a single SAR image, a stack of reconstructed images is generated. These images are then merged using a gradient-aware 3D filtering process that distinguishes between edge and non-edge regions. Experimental results on TerraSAR-X data demonstrate that this approach significantly improves visual quality and quantitative metrics such as the Equivalent Number of Looks (ENL) and Edge Save Index (ESI) compared to state-of-the-art methods.
Introduction
Synthetic Aperture Radar (SAR) provides critical high-resolution Earth observation data regardless of weather or light conditions. However, the coherent nature of SAR imaging introduces "speckle," a multiplicative noise that degrades image interpretability and complicates automated analysis. Traditional denoising methods often struggle to balance noise suppression with the preservation of fine spatial details. This paper proposes a framework that synthesizes multiple observations from a single image to exploit statistical redundancy for superior despeckling.
Proposed Methodology
The proposed framework operates in three primary stages: subset selection, multiple compressive reconstruction, and selective 3D filtering.
1. Subset Selection
To simulate multiple observations from a single source, the original SAR image is partitioned into $P$ mutually exclusive pixel groups. From these, $T$ distinct subsets are formed. A key constraint is that each subset must contain at least 20% unique pixels compared to any other subset, ensuring diverse noise realizations across the generated "views."
2. Multiple Compressive Reconstruction
Each subset serves as a measurement vector within a Compressive Sensing framework. Using the GPSR algorithm and Haar wavelets, a full SAR image is reconstructed for each subset. The inherent sparsity constraints of CS reconstruction act as a preliminary noise filter, as high-frequency speckle is less likely to be represented in the sparse wavelet domain.
3. Selective 3D Filtering
The final "clean" image is produced by merging the stack of reconstructed images (70 in the implemented study). A gradient-based decision mechanism determines the filtering strategy for each pixel to optimize the balance between smoothing and detail preservation:
- Edge Pixels: In regions with high gradients (features), a 1x1xT mean kernel (pixel-wise average) is used to maintain sharp transitions.
- Non-Edge Pixels: In smooth regions with low gradients, a 3x3xT median filter is applied to aggressively suppress noise across both spatial and temporal (stack) dimensions.
Experimental Results
The technique was validated using real-world TerraSAR-X satellite data across maritime, urban, and rural scenes.
Quantitative Performance
The proposed method consistently outperformed benchmarks such as Lee, Frost, BM3D, and PPB. Notably, for complex urban scenes in Tianjin, the method achieved an Equivalent Number of Looks (ENL) of 44.96, nearly doubling the 26.22 achieved by the high-performing BM3D algorithm. High Edge Save Index (ESI) values further confirmed that the gradient-aware filtering successfully protected structural details while smoothing homogeneous areas.
Visual Quality
Qualitative analysis shows that the proposed method produces cleaner images with fewer artifacts than patch-based methods. The "synthetic multi-looking" effectively smooths background noise, while the selective filter prevents the blurring of critical edges, such as those found on ships or urban infrastructure.
Conclusion
By combining Compressive Sensing with a selective 3D filtering strategy, this research provides a robust solution for SAR image despeckling. The ability to generate and statistically merge multiple reconstructions from a single source image overcomes the limitations of traditional single-image filters, offering a significant advancement for high-resolution radar imaging applications.
Summary updated