Heuristic Quality Coefficients for Interferometric Phase Linking

1. Problem Statement

In Multitemporal Interferometric Synthetic Aperture Radar (InSAR), Distributed Scatterers (DS) are prevalent in vegetated and non-urban areas. Unlike Persistent Scatterers (PS), DS targets suffer from temporal decorrelation, leading to inconsistent interferometric phases across acquisition pairs. To retrieve surface deformation, Phase Linking (PL) methods are used to estimate a single-reference phase history from the sample coherence matrix.

However, a critical practical challenge exists: reliability assessment.

• While Maximum-Likelihood (ML) estimators allow for uncertainty quantification via Fisher information, the broad range of non-ML PL methods (e.g., Phase Triangulation, Eigendecomposition) lacks generally applicable, standardized uncertainty metrics.
• Existing reliability indicators are often specific to a single algorithm, depend heavily on the number of acquisitions, and lack a unified scale, making it difficult to filter unreliable pixels for downstream deformation analysis.

2. Methodology and Theoretical Framework

The authors propose a unified mathematical framework to categorize common PL methods and derive three new, normalized quality coefficients.

A. Unified Optimization Framework

The paper models PL as a (weighted) rank-one approximation problem. Given a sample coherence matrix $T = |T| \circ \Phi$ (where $\Phi$ is the unit phasor matrix), PL methods seek a phase history vector $\hat{\theta}$ that maximizes the alignment with the observed phases.
The general estimator form is:
$$\hat{\Theta}_R = \{ \angle v \mid v \in \arg\max_{v \in X} \langle vv^H, W \circ \Phi \rangle_F \}$$
Where:

• $X$ is the feasible set: $T_N$ (unit modulus vectors) for Phase Triangulation (PT) or $U_N$ (unit norm vectors) for Eigendecomposition (ED).
• $W$ is a weight matrix (e.g., Equal-Weighted, Coherence-Weighted, or Maximum-Likelihood).
• This framework encompasses methods like SBAS, CAESAR, EMI, and LaMIE.

B. Three Heuristic Quality Coefficients

The authors introduce three scalar coefficients, normalized to the interval $[0, 1]$ (where 1 is perfect reliability and 0 is pure noise behavior).

1. Goodness-of-Fit Coefficient ($\gamma_{GOF}$)

• Concept: Measures how well the estimated phase history fits the observed data relative to a theoretical upper bound (perfect coherence) and a modeled noise floor.
• Mechanism: It normalizes the achieved objective function value $f(T, \hat{\theta})$ using:
  • $f_{max}(T)$: The theoretical maximum (ideal case).
  • $f_{min}(T)$: A lower bound derived from the expected objective value under pure noise ($\Sigma = I_N$).
• Noise Floor Modeling: Since $f_{min}$ is hard to derive analytically, the authors use Monte Carlo simulations to fit a logistic sigmoid model based on stack size ($N$) and sample size ($W$).
• Formula: $\gamma_{GOF} = \max\left( \frac{F(T, \hat{\theta}) - F_{noise}}{1 - F_{noise}}, 0 \right)$.

2. Closure Phase Coefficient ($\gamma_{CP}$)

• Concept: An a priori indicator computed directly from the sample coherence matrix before running the PL solver. It quantifies internal phase consistency.
• Mechanism: Based on the closure phase $\phi^\Delta_{ijk} = \angle(\Phi_{ij}\Phi_{jk}\Phi_{ki})$. In a perfect phase history, the sum of phases around a triangle is zero. Non-zero closure phases indicate noise or inconsistent scattering.
• Variants:
  • Equal-Weighted ($\gamma_{CP}$): Simple average of cosines of closure phases.
  • Coherence-Weighted ($\gamma_{CPw}$): Weights triplets by coherence magnitude, making it more robust for ML/CW methods.
• Utility: Serves as a fast pre-screening tool to filter unreliable pixels before expensive optimization.

3. Ambiguity Coefficient ($\gamma_{A}$)

• Concept: Measures the uniqueness of the solution. It assesses whether the primary solution is significantly better than the best alternative in its orthogonal complement.
• Mechanism:
  1. Find the primary estimate $\hat{\theta}^{(1)}$.
  2. Find the best secondary estimate $\hat{\theta}^{(2)}$ constrained to be orthogonal to $\hat{\theta}^{(1)}$ in the complex inner product space.
  3. Compare the objective values: $\gamma_A$ is high if the primary solution is much better than the secondary; it is low if multiple distinct solutions explain the data equally well (ambiguity).
• Implementation: For ED methods, this is computationally cheap (using the second largest eigenvalue). For PT methods, it requires solving a constrained nonlinear optimization problem (using a robust Polak-Ribière Conjugate Gradient solver).

3. Key Contributions

1. Unified Categorization: Established a common optimization form for diverse PL methods (PT vs. ED, various weighting schemes), clarifying their mathematical relationships.
2. Standardized Metrics: Introduced three normalized coefficients ($\gamma_{GOF}, \gamma_{CP}, \gamma_{A}$) that are comparable across different methods and dataset sizes.
3. Noise Floor Modeling: Developed an empirical model to estimate the noise floor for the Goodness-of-Fit coefficient, enabling consistent normalization across varying stack sizes and sample counts.
4. Robust Solvers: Proposed a penalized Polak-Ribière Conjugate Gradient (PRCG) approach to efficiently compute the Ambiguity coefficient for Phase Triangulation methods.

4. Experimental Results

The authors evaluated the coefficients using:

• Simulations: Exponential and seasonal decorrelation models with varying noise levels.
• Real Data: A TerraSAR-X stack (44 acquisitions) over Visp, Switzerland (urban, agricultural, and forested areas).

Simulation Findings:

• Goodness-of-Fit ($\gamma_{GOF}$): Showed the strongest and most consistent correlation with the normalized absolute phase error ($nAE$) across all methods and decorrelation scenarios. It is the most reliable default indicator.
• Closure Phase ($\gamma_{CP}$): Provided a stable predictor, particularly the coherence-weighted version ($\gamma_{CPw}$). It is highly effective for pre-screening.
• Ambiguity ($\gamma_{A}$): Showed mixed results for ML methods but strong positive correlation for Equal-Weighted (EW) and Coherence-Weighted (CW) methods. It effectively identifies cases where the solution is not unique.

Real-Data Findings (Visp, Switzerland):

• Spatial Patterns: High coefficients were observed in stable urban areas and well-maintained fields (e.g., a stadium). Low coefficients appeared in rivers and dense vegetation.
• Land Cover Change: The Ambiguity coefficient successfully highlighted areas with temporal changes (e.g., new construction). In these areas, the primary and secondary solutions diverged, lowering $\gamma_A$ even when $\gamma_{GOF}$ was moderate.
• Complementarity: The three coefficients provide distinct information:
  • $\gamma_{CP}$: Fast check for internal consistency.
  • $\gamma_{GOF}$: Measure of model fit.
  • $\gamma_{A}$: Measure of solution uniqueness/stability.

5. Significance and Conclusion

This work addresses a critical gap in DS-InSAR processing by providing generalizable, normalized reliability indicators for non-ML phase linking methods.

• Operational Impact: The coefficients allow for automated, data-driven filtering of pixels. $\gamma_{CP}$ can be used for fast pre-screening, while $\gamma_{GOF}$ and $\gamma_{A}$ can be used to select the most reliable pixels for deformation analysis.
• Diagnostic Power: The Ambiguity coefficient is particularly novel, offering a way to detect "hidden" failures where the model fits the data well but the solution is physically ambiguous (e.g., due to changing scattering mechanisms).
• Future Work: The framework is currently limited to estimators fitting the rank-one approximation form. Future work could extend this to SBAS explicitly and validate transferability across different sensors (e.g., Sentinel-1, ALOS-2).

In summary, the paper provides a robust toolkit for improving the quality control of distributed scatterer InSAR processing, moving beyond method-specific heuristics to a unified, mathematically grounded approach.