TL;DR
This paper introduces a novel multibaseline InSAR phase denoising method that leverages the statistical property of kurtosis. By proposing a "baseline-invariant property of kurtosis," the authors develop a more accurate noise variance estimation technique, leading to improved filtering performance compared to existing single-baseline methods.
What Problem Does This Paper Address?
Synthetic Aperture Radar (SAR) interferometry (InSAR) is a powerful technique for mapping topography and monitoring surface deformation. Multi-baseline InSAR, which uses multiple interferograms with varying baselines, offers advantages in handling complex terrain and improving accuracy. However, a critical step in InSAR processing is phase filtering to remove noise. Existing multi-baseline filtering methods often adapt single-baseline techniques, failing to fully exploit the benefits of multiple baselines. This paper addresses the challenge of effectively denoising multi-baseline InSAR data by developing a statistically-driven approach that utilizes the unique properties of interferometric phase.
Methods
The proposed method consists of four main parts:
- Non-Subsampled Shearlet Transform (NSST) Decomposition: The interferograms are decomposed into multi-scaled and multi-directional coefficients using NSST. This transform is chosen for its ability to effectively represent the fringe structures and distributed discontinuities present in interferograms.
- Noise Level Estimation based on Kurtosis: This is the core innovation. The authors introduce the concept of the "baseline-invariant property of kurtosis," an empirical observation that the kurtosis of interferograms with different baselines remains relatively constant. This property, combined with a kurtosis model, allows for a more accurate estimation of noise variance. This estimation is formulated as a constrained non-convex optimization problem.
- Pre-thresholded Wiener Filter: A Wiener filter is applied to the NSST coefficients. This filter uses the estimated noise variance to shrink or zero out noisy coefficients while preserving important signal information.
- Inverse NSST: The filtered NSST coefficients are transformed back to the spatial domain to obtain the denoised interferograms.
Datasets
The paper uses simulated datasets generated from various topographic models (cone, building, plain, basin, plateau, steep mountain) and real InSAR data from the Alos-1 satellite over the Colorado Grand Canyon.
Key Insights and Ideas
- Baseline-Invariant Property of Kurtosis: The observation that kurtosis remains constant across different baselines for the same scene is a crucial insight. This statistical property is key to overcoming the limitations of traditional noise estimation methods.
- Kurtosis as a Noise Indicator: Kurtosis, a measure of the "tailedness" of a probability distribution, proves to be a sensitive indicator of noise in interferograms. Higher kurtosis generally correlates with cleaner signals.
- Leveraging Multi-baseline Data: The method effectively utilizes the redundancy and complementary information from multiple baselines for more robust noise estimation, a significant improvement over single-baseline approaches.
- NSST for Fringe Representation: The NSST transform is well-suited for capturing the directional and multi-scale nature of interferometric fringes, making it an effective domain for applying statistical analysis and filtering.
Key Results
- Improved Noise Variance Estimation: The proposed kurtosis-based noise level estimation method demonstrates higher accuracy and stability compared to traditional methods, especially in low coherence regions.
- Superior Filtering Performance: Experiments on simulated and real data show that the proposed NSST-based method outperforms several state-of-the-art single-baseline filters (Goldstein, LFE, OADF, NL-InSAR, InSAR-BM3D) in terms of Mean Square Error (MSE) and Gradient Magnitude Similarity (GMS).
- Edge Preservation: The method effectively preserves phase gradients and edges, which is critical for subsequent InSAR processing steps like phase unwrapping.
- Efficiency: While not the fastest, the proposed method offers a good balance between performance and computation time, outperforming some advanced methods in efficiency. For example, it achieved a computation time of 12.68 seconds on a cone interferogram, second only to the Goldstein method.
Limitations and Future Work
- Memory Requirements: The NSST decomposition can occupy a significant amount of memory due to the large number of coefficient components, especially with a high number of scales. The authors suggest this can be managed by adjusting the NSST decomposition scale based on available computational resources.
- Complex Topography: While the method performs well, the paper notes that recovering the real phase correctly at very complex topography regions can still be challenging, potentially due to inherent limitations of interferograms with very long baselines.
- Future Work: The paper implicitly suggests further exploration of statistical properties of InSAR data and their application to denoising and other InSAR processing tasks. The authors also highlight the potential for further optimization of the computational efficiency.
Summary updated