main.pdf
Introduction
In multitemporal interferometric synthetic aperture radar (InSAR), phase estimation and phase linking over distributed scatterers rely on averaging observations from selected neighboring pixels ((Ho Tong Minh and Tebaldini, 2023)). This raises three questions for each pixel to be estimated, namely whether an average over its neighbors can represent it at all, which neighbors are statistically homogeneous with it, and how large the set of averaged pixels should be. We call the pixel to be estimated the center pixel, or center, and the set of averaged pixels the ensemble. The first question arises in particular for pixels in layover and shadow areas, which are typically masked. In layover, scatterers at several heights are projected onto the same pixel in the synthetic aperture radar (SAR) image, so that subtracting a topographic phase may introduce systematic phase inconsistencies. In shadow, the phase is often dominated by noise that is less correlated between neighboring pixels than speckle. Geometric layover and shadow maps derived from the acquisition geometry and a digital elevation model (DEM) (Kropatsch and Strobl, 1990) can identify these areas, but they depend on the accuracy of the DEM. Layover and shadow caused by objects that are not included in the model, such as buildings and trees, are another limitation. Layover has also been detected from the data itself. Tupin et al. (1996) use the amplitude, Zhu et al. (2017) the local fringe frequency of an interferogram over high-rise buildings, and Rossi and Eineder (2015) the number of times each SAR pixel occurs in the geocoding of an elevation model generated from the interferometric data, without external data. Tupin et al. (1996) also detect areas without topographic phase from the filtered coherence and the confidence of the local fringe frequency estimates of an interferogram. These areas include shadow, decorrelated areas and surfaces without backscatter.
Established processing frameworks such as SqueeSAR [3 ] identify statistically homogeneous pixels (SHPs) by tests on the temporal amplitude distributions [14 ] (Even and Schulz 2018). Later tests also use the interferometric phases [24 ] (Dong et al. 2018; Song et al. 2017). Other approaches collect pixels of similar intensity by region growing (Vasile et al., 2006) or weight pixels by the similarity of their surrounding patches (Deledalle et al., 2011, 2015). SHP tests that compare the temporal mean amplitudes or intensities of two pixels lose their significance level when the speckle is temporally correlated. On a coherent target the speckle changes little between acquisitions, so that the means of two pixels of the same surface keep the difference of their speckle. A threshold set for independent samples or for one assumed coherence therefore holds only at that coherence.
Averaging reduces noise, but its extent determines how much spatial detail is retained. Strongly decorrelated targets need larger ensembles to reach a given precision (cf. Rocca, 2007, Eqs. (5.22) to (5.25)). Where the phase varies rapidly in space, as in areas with strong spatial gradients of line-of-sight deformation, a smaller ensemble may suffice if the coherence is high enough, whereas a large one blurs fringes and lowers the estimated coherence (Zebker and Chen, 2005). The averaging neighborhood has been adapted to this in shape and size. For single interferograms, directional windows follow the local fringe orientation (Lee et al., 1998), and the intersection of confidence intervals (ICI) rule chooses, for each pixel, the largest window whose local polynomial estimate stays consistent with those of smaller windows [8 ] (Katkovnik et al. 2002). For stacks, Pepe et al. [15 ] adapt a fixed multilook window to the phase by removing the pixels whose phases deviate from the multilooked phases of that window and comparing the deviations of the remaining pixels with those of the pixel that deviates least. In other SHP-based processing, however, the coherence matrix is usually estimated from all pixels that the SHP selection accepts (e.g., [6 , 24 ] ), also where the phase varies within the search window. Chi et al. (2022, Section 3.2 and Eq. (13)) instead shrink the search window of an SHP selection at each pixel until a coefficient of variation of the phase of the time-averaged differential interferogram falls below a threshold, which requires unwrapped interferograms.
In this paper, we examine a simple quantity, the squared magnitude of the mean phasor of the relative phase of two pixels across acquisitions, which we refer to as spatial phase coherence (SPC), and its use for all three questions. The quantity itself is known under several names (Section 2.2). It is used in directional statistics (Mardia and Jupp, 2000) and in InSAR for persistent scatterers (Ferretti et al., 2000), and in related forms for phase stability (Hooper et al., 2004, 2007), point coherence estimation (Vecchioli et al., 2023) and SHP selection [13 , 15 , 17 ] . Less attention has so far been paid to the behavior of the statistic on single-look data with spatially correlated speckle, and to the uses of its properties in the processing chain, from masking to the size of the ensemble that an SHP selection returns. Our contributions are as follows.
A statistical model of SPC on single-look data, with (a) its expectation under the circular complex Gaussian (CCG) model, (b) its floor under speckle that adjacent pixels share through the system point
spread function, predicted from the processing parameters, and (c) an identity that links pairwise SPC to the mean squared phase-only coherence of an ensemble.
Two maps, the neighborhood SPC (NSPC) and the phase coherence indicator (PCI), computed from fixed sets of neighbors at the cost of only one inner product over the acquisitions per neighbor, and available directly from the single-look complex (SLC) stack. A threshold on NSPC, the average SPC of a pixel with its eight direct neighbors, flags pixels where this average lies below the floor that shared speckle sets, and complements geometric layover and shadow maps. The PCI, an average of SPC over neighbors that share no speckle with the pixel, rises with its coherence and steers the SHP selection and the sizing of the ensemble, as described below.
Thresholds for SHP tests that compare temporal mean amplitudes or intensities. They depend on the PCI and keep the false rejection rate of the order of the significance level at every coherence, also where the speckle of single-look data is temporally correlated.
Adaptive neighborhood sizing (ANS) in two parts: a search window scaled with the PCI (S), and a truncation step (T) that reduces a selected ensemble to a smaller nested one only where that one agrees clearly better with the phases of the center and of the pixels that share its speckle. The agreement is measured with a generalization of SPC, and ANS can be combined with any SHP selection.
The two maps, the thresholds and ANS are evaluated in simulations with known phase and demonstrated on TerraSAR-X data over Visp, Switzerland, and Sentinel-1 data over Gilgit-Baltistan, Pakistan. ANS is combined with three kinds of SHP selection: (a) the amplitude-based fast SHP selection (FaSHPS) of [6 ] , which tests the temporal amplitudes of two pixels, (b) the intensity- and phase-based covariance matrix patch (CMP) selection of [24 ] and (c), as their reference, all pixels of the search window (the full window). In § 2.1 , we develop the theory, and in § 2.2 , we define the mask, the PCI, the SHP selections, the two parts of ANS and the estimator of the coherence matrix. In § 2.3 and § 3 , we present the simulations and the real data, including the runtime. In § 3.2 , we discuss the applicability of the two maps and of ANS, the variants not adopted, and limitations.
Spatial phase coherence
Definition
Consider a stack of N > 1 N > 1 N > 1 coregistered SLC images in radar coordinates, and let x p ( n ) x_p(n) x p ( n ) denote the complex value of pixel p p p in acquisition n n n . A pixel is identified with its integer image coordinates (azimuth, range), so that q − p q-p q − p denotes the spatial lag from pixel p p p to pixel q q q . A model phase, such as a simulated topographic phase, may have been subtracted beforehand (§ 2.1.6 ). For nonzero observations, define
u p ( n ) = x p ( n ) ∣ x p ( n ) ∣ = e i ϕ p ( n ) , u p = ( u p ( 1 ) , … , u p ( N ) ) T . (1) u_p(n) = \frac{x_p(n)}{\lvert x_p(n)\rvert} = e^{i\phi_p(n)}, \qquad\mathbf{u}_p = \left(u_p(1),\ldots,u_p(N)\right)^{\mathsf{T}}. \tag*{(1)} u p ( n ) = ∣ x p ( n )∣ x p ( n ) = e i ϕ p ( n ) , u p = ( u p ( 1 ) , … , u p ( N ) ) T . (1)
The coherence of two acquisitions s , t s,t s , t is the normalized covariance of their complex values (Even and Schulz, 2018, Eq. (2) ). It is estimated by the sample coherence over a neighborhood W \mathcal{W} W (Even and Schulz, 2018, Eq. (8) ), whose phase-only counterpart is
γ ^ s t ∞ ( W ) = ∣ 1 ∣ W ∣ ∑ q ∈ W u q ( s ) u q ( t ) ‾ ∣ . (2) \widehat{\gamma}_{st}^{\infty}(\mathcal{W}) = \left\lvert\frac{1}{\lvert\mathcal{W}\rvert} \sum_{q\in\mathcal{W}} u_q(s)\overline{u_q(t)} \right\rvert. \tag*{(2)} γ s t ∞ ( W ) = ∣ W ∣ 1 q ∈ W ∑ u q ( s ) u q ( t ) . (2)
Exchanging the roles of pixels and acquisitions in Eq. (2) , i.e., fixing two pixels p , q p,q p , q and averaging over the acquisitions (Figure 1 ), gives the SPC of the two pixels, which we define as the square of the resulting magnitude,
Figure 1. Illustration of phase-only coherence and SPC on a stack of images. (a) The phase-only coherence (Eq. (2) ) of two acquisitions s s s , t t t is built from the sum of the products u q ( s ) u q ( t ) u_q(s)u_q(t) u q ( s ) u q ( t ) over the pixels q q q of a neighborhood W \mathcal{W} W , the same in both images. (b) SPC (Eq. (3) ) of two pixels p p p , q q q is built from the sum of the products u p ( n ) u q ( n ) u_p(n)u_q(n) u p ( n ) u q ( n ) over the N N N acquisitions of the stack.
S p q = ∣ 1 N ∑ n = 1 N u p ( n ) ‾ u q ( n ) ∣ 2 = ∣ u p H u q ∣ 2 N 2 . (3) S_{pq} = \left\lvert\frac{1}{N}\sum_{n=1}^{N}\overline{u_p(n)}u_q(n) \right\rvert^2 = \frac{\left\lvert\mathbf{u}_p^{\mathsf{H}}\mathbf{u}_q \right\rvert^2}{N^2}. \tag*{(3)} S pq = N 1 n = 1 ∑ N u p ( n ) u q ( n ) 2 = N 2 u p H u q 2 . (3)
It satisfies 0 ≤ S p q ≤ 1 0 \le S_{pq} \le1 0 ≤ S pq ≤ 1 , with S p q = 1 S_{pq}=1 S pq = 1 precisely when the two phase vectors differ by a constant offset. SPC is invariant to constant phase rotations of either pixel and to any acquisition-dependent rotation common to both. Referencing all phases to one acquisition is a rotation of the first kind and leaves SPC unchanged, provided that the reference acquisition, whose phase is then zero at every pixel, is kept as one of the N N N terms of the sum. Like Eq. (2) , SPC uses the phases only, as do the estimates of the coherence matrix in this paper (§ 2.2.8 ). Exchanging the roles in the sample coherence instead would give the correlation coefficient of the two complex time series. Bright samples and acquisitions would dominate it, and its expectation, the floor of § 2.1.5 and the thresholds derived from it would depend on amplitude changes over the acquisitions, e.g., from soil moisture or vegetation. SPC does not change under any positive scaling of the samples, and the amplitudes enter the processing only through the SHP tests (§ 2.2.4 and § 2.2.5 ).
If the relative phases ϕ q ( n ) − ϕ p ( n ) \phi_q(n)-\phi_p(n) ϕ q ( n ) − ϕ p ( n ) are independent across the acquisitions and uniformly distributed, E [ S p q ] = 1 / N \mathbb{E}[S_{pq}]=1/N E [ S pq ] = 1/ N . Removing this bias and rescaling gives the corrected SPC and its clipped version
Q p q = N S p q − 1 N − 1 , Q p q + = max ( Q p q , 0 ) , (4) Q_{pq}=\frac{NS_{pq}-1}{N-1}, \qquad Q_{pq}^{+}=\max(Q_{pq},0), \tag*{(4)} Q pq = N − 1 N S pq − 1 , Q pq + = max ( Q pq , 0 ) , (4)
respectively. Where the distinction does not matter, we also call Q p q Q_{pq} Q pq , Q p q + Q_{pq}^{+} Q pq + and their square roots SPC.
Relation to existing measures
The quantities in Eqs. (3) and (4) are known in InSAR and in other scientific fields under different names. Section S1.1 of the supplement gives the details of the relations below. In directional statistics, testing S p q S_{pq} S pq against a threshold is the Rayleigh test of uniformity, applied here to the relative phases ((Mardia and Jupp, 2000, Section 6.3.1, Eqs. (6.3.3) and (6.3.4)). In InSAR, with the phase model of the multi-image coherence set to zero, S p q \sqrt{S_{pq}} S pq has the form of the multi-image phase coherence of a pair of neighboring persistent scatterers p p p and q q q ((Ferretti et al., 2000, Eq. (2.7))), summed over the acquisitions instead of single-reference interferograms. Vecchioli et al. (2023) estimate the coherence of each point of a stack from such coherences of pairs of nearby points, computed after a motion model is fitted to each pair. They assume that the phase noise of two points is independent, which may require leaving out the nearest pixels of oversampled images. The expected coherence of a pair is then the product of the coherences of its two points, and the logarithms of these products form an overdetermined linear system for the coherence of every point. The phase-stability measure of Hooper et al. (2007, Eq. (20)) applies the same construction, the magnitude of a mean phasor, to the phases of one pixel relative to the filtered phase of the surrounding pixels, after a fitted look angle error is removed. For SHP selection, the zero-lag phase cross-correlation of Narayan et al. [13 ] has magnitude N I S p q N_I\sqrt{S_{pq}} N I S pq when S p q S_{pq} S pq is computed over the N I N_I N I single-reference interferograms that it uses, instead of over the acquisitions. The magnitude of the centered complex correlation of Shi et al. [17 ] Eq. (1)) differs from $\sqrt{S_{pq}}$ in general and coincides with it when both temporal phasor means vanish. The circular variance used for adaptive multilooking by {citet}Pepe et al. (2021, Eqs. (9) and (10) ) is one minus the magnitude of a mean phasor, taken over the interferograms and computed between the interferometric phases of a pixel and the multilooked phases of the window rather than those of a second pixel. The SPC formula is therefore not a contribution of this paper. The contributions concern its behavior on single-look SAR data and its use for masking, for an indicator of local coherence that steers the SHP selection and for the size of an ensemble. Its statistics have also been examined in neuroscience: Q p q Q_{pq} Q pq is the pairwise phase consistency of Vinck et al. [21 ] Eqs. (14) and (15), an unbiased estimator of squared phase synchrony between neuronal signals recorded over repeated trials, and $\sqrt{S_{pq}}$ is the phase-locking value of {citet}Lachaux et al. (1999, p. 195). Acquisitions take the place of trials here. The unbiasedness of the pairwise phase consistency assumes independent, identically distributed trials. Acquisitions are neither, and $\mathbb{E}[Q_{pq}]$ is then the mean expected cosine agreement ({xref}Eq. (5)`).
Agreement across interferograms and coherence of an ensemble
Phase linking uses the coherence of an ensemble, whereas SPC compares pixels pairwise. On average the two agree: an ensemble is coherent over the interferograms exactly when its pixels agree in phase pairwise over the acquisitions, and the coherence of an ensemble is maximal in every interferogram exactly when the SPC of each pair of its pixels is maximal. Write Δ ϕ p s t = ϕ p ( s ) − ϕ p ( t ) \Delta\phi_{p}^{st}=\phi_p(s)-\phi_p(t) Δ ϕ p s t = ϕ p ( s ) − ϕ p ( t ) , v p s t = e i Δ ϕ p s t v_p^{st}=e^{i\Delta\phi_p^{st}} v p s t = e i Δ ϕ p s t and M = ( N 2 ) M=\binom{N}{2} M = ( 2 N ) . Expanding Eq. (3) gives
Q p q = 1 M ∑ s < t cos ( Δ ϕ p s t − Δ ϕ q s t ) . (5) Q_{pq}=\frac{1}{M}\sum_{s<t}\cos\left(\Delta\phi_p^{st}-\Delta\phi_q^{st}\right). \tag*{(5)} Q pq = M 1 s < t ∑ cos ( Δ ϕ p s t − Δ ϕ q s t ) . (5)
Corrected SPC is thus the mean cosine of the difference between the interferometric phases of the two pixels over all M M M interferograms (§ 13.1.1 of the supplement), computed in O ( N ) O(N) O ( N ) operations without forming the interferograms or a coherence matrix. The identity with the coherence of an ensemble holds for any set E \mathcal{E} E of L = ∣ E ∣ ≥ 2 L=|\mathcal{E}|\ge2 L = ∣ E ∣ ≥ 2 pixels,
1 M ∑ s < t L γ ~ s t ( E ) 2 − 1 L − 1 = Q ˉ ( E ) = 1 L ( L − 1 ) ∑ q , q ′ ∈ E q ≠ q ′ Q q q ′ . (6) \frac{1}{M}\sum_{s<t}\frac{L\tilde{\gamma}_{st}(\mathcal{E})^2-1}{L-1}=\bar{Q}(\mathcal{E})=\frac{1}{L(L-1)}\sum_{\substack{q,q'\in\mathcal{E}\\q\ne q'}}Q_{qq'}. \tag*{(6)} M 1 s < t ∑ L − 1 L γ ~ s t ( E ) 2 − 1 = Q ˉ ( E ) = L ( L − 1 ) 1 q , q ′ ∈ E q = q ′ ∑ Q q q ′ . (6)
The left side corrects the squared coherence magnitude of each interferogram for its bias 1 / L 1/L 1/ L at zero coherence for L L L independent looks (cf. Touzi et al. (1999, Eq. (5)), as Eq. (4) corrects S p q S_{pq} S pq . The mean corrected SPC over all pairs of an ensemble is thus the mean corrected squared phase-only coherence of that ensemble (§ 13.1.1 of the supplement). Unlike this mean over all pairs, which includes pairs that do not contain the center, the maps of § 2.2.1 and § 2.2.2 are built from the pairs of each pixel with fixed neighbors and describe that pixel and its immediate surroundings. The two maps, NSPC and the PCI, need eight and 24 inner products of length N N N per pixel, against O ( L N 2 ) O(LN^2) O ( L N 2 ) operations for the coherence of an ensemble of L L L pixels.
Expected value under the CCG model
The expected value of SPC shows how much of a low SPC comes from decorrelation and how much from different true phase histories. It follows from the expectation of each cosine in Eq. (5) , which depends on the distribution of the single-look interferometric phase Δ ϕ \Delta\phi Δ ϕ of an interferogram with coherence γ \gamma γ . Let g g g denote the mean cosine of its phase error Δ ϕ − arg γ \Delta\phi-\arg\gamma Δ ϕ − arg γ (cf. Just and Bamler, [7 ] Eq. (15); {citet}Lee et al., 1994, Eq. (21) ), in the closed form of {citet}López-Martínez and Fàbregas (2002, Eq. (8) )`, with 2 F 1 _2F_1 2 F 1 the Gauss hypergeometric function,
g ( ∣ γ ∣ ) = E [ cos ( Δ ϕ − arg γ ) ] = π 4 ∣ γ ∣ 2 F 1 ( 1 2 , 1 2 ; 2 ; ∣ γ ∣ 2 ) . (7) g(|\gamma|)=\mathbb{E}[\cos(\Delta\phi-\arg\gamma)]=\frac{\pi}{4}|\gamma|\,{}_2F_1\left(\frac{1}{2},\frac{1}{2};2;|\gamma|^2\right). \tag*{(7)} g ( ∣ γ ∣ ) = E [ cos ( Δ ϕ − arg γ )] = 4 π ∣ γ ∣ 2 F 1 ( 2 1 , 2 1 ; 2 ; ∣ γ ∣ 2 ) . (7)
For example, g ( 0.2 ) = 0.16 g(0.2) = 0.16 g ( 0.2 ) = 0.16 , g ( 0.5 ) = 0.41 g(0.5) = 0.41 g ( 0.5 ) = 0.41 , g ( 0.8 ) = 0.70 g(0.8) = 0.70 g ( 0.8 ) = 0.70 . Suppose pixels p p p and q q q are independent and each follows a zero-mean CCG distribution with true coherence matrix entries γ s t p \gamma_{st}^{p} γ s t p , γ s t q \gamma_{st}^{q} γ s t q (not sample estimates) and true phase histories θ p \theta_p θ p , θ q \theta_q θ q , with arg γ s t p = Δ θ p s t = θ p ( s ) − θ p ( t ) \operatorname{arg}\gamma_{st}^{p} = \Delta\theta_{p}^{st} = \theta_p(s) - \theta_p(t) arg γ s t p = Δ θ p s t = θ p ( s ) − θ p ( t ) . Since p p p and q q q are independent and the phase error of a CCG pixel is symmetric about zero (Section S1.2 of the supplement), taking expectations in Eq. (5) gives
E [ Q p q ] = 1 M ∑ s < t g ( ∣ γ s t p ∣ ) g ( ∣ γ s t q ∣ ) cos ( Δ θ p s t − Δ θ q s t ) . (8) \mathbb{E}[Q_{pq}] = \frac{1}{M}\sum_{s<t} g\left(\lvert\gamma_{st}^{p}\rvert\right)g\left(\lvert\gamma_{st}^{q}\rvert\right)\cos\left(\Delta\theta_{p}^{st}-\Delta\theta_{q}^{st}\right). \tag*{(8)} E [ Q pq ] = M 1 s < t ∑ g ( ∣ γ s t p ∣ ) g ( ∣ γ s t q ∣ ) cos ( Δ θ p s t − Δ θ q s t ) . (8)
Decorrelation lowers E [ Q p q ] \mathbb{E}[Q_{pq}] E [ Q pq ] through g 2 g^2 g 2 , and differences in phase history through the cosine. Even two pixels with the same true phase history and ∣ γ s t ∣ = 0.5 \lvert\gamma_{st}\rvert= 0.5 ∣ γ s t ∣ = 0.5 for all acquisition pairs give E [ Q p q ] ≈ 0.17 \mathbb{E}[Q_{pq}] \approx0.17 E [ Q pq ] ≈ 0.17 , so that where the coherence is low, a single value of SPC says little about whether two pixels share a true phase history.
When the relative phasors z n = u p ( n ) ‾ u q ( n ) z_n = \overline{u_p(n)}u_q(n) z n = u p ( n ) u q ( n ) are independent across acquisitions with means ξ n \xi_n ξ n , the expectation (Section S1.2 of the supplement) is
E [ Q p q ] = ∣ ∑ n ξ n ∣ 2 − ∑ n ∣ ξ n ∣ 2 N ( N − 1 ) , (9) \mathbb{E}[Q_{pq}] = \frac{\left\lvert\sum_n \xi_n\right\rvert^2-\sum_n\lvert\xi_n\rvert^2}{N(N-1)}, \tag*{(9)} E [ Q pq ] = N ( N − 1 ) ∣ ∑ n ξ n ∣ 2 − ∑ n ∣ ξ n ∣ 2 , (9)
which reduces to ∣ ξ ∣ 2 \lvert\xi\rvert^2 ∣ ξ ∣ 2 when all ξ n \xi_n ξ n equal a common mean ξ \xi ξ (cf. [21 Eqs. (13) and (19] ). We use Eq. (9) for the null hypothesis below.
SPC under correlated speckle
If the relative phases are independent across the acquisitions and uniformly distributed, E [ Q p q ] = 0 \mathbb{E}[Q_{pq}] = 0 E [ Q pq ] = 0 . This does not hold for nearby pixels in SLC data. The SAR point spread function spreads each scatterer over several samples, so that neighboring pixels share speckle. Let ρ ℓ \rho_\ell ρ ℓ denote the complex correlation of the speckle at spatial lag ℓ \ell ℓ . Under the null hypothesis that the pixels carry no common signal beyond this shared speckle, and if the observations are independent across acquisitions, the relative phasors have mean magnitude g ( ∣ ρ ℓ ∣ ) g(\lvert\rho_\ell\rvert) g (∣ ρ ℓ ∣) (Section S1.2 of the supplement), and Eq. (9) gives
E ρ [ Q p q ] = g ( ∣ ρ ℓ ∣ ) 2 . (10) \mathbb{E}_{\rho}[Q_{pq}] = g(\lvert\rho_\ell\rvert)^2. \tag*{(10)} E ρ [ Q pq ] = g (∣ ρ ℓ ∣ ) 2 . (10)
We call the level that SPC reaches under this null hypothesis the correlated-noise floor. We refer to distributed scattering without point scatterers as clutter, the name that persistent scatterer processing gives to the background of a stable scatterer ((Ferretti et al., 2001, p. 11)), and to clutter under this null hypothesis as zero-coherence clutter. We predict the correlation from the processing parameters of the SLC data and check the prediction with an estimate from the stack itself (Section 5.2). For a spectral weighting W ( f ) W(f) W ( f ) over the processed bandwidth B B B and a sampling rate f s f_s f s , the speckle correlation at a lag of ℓ \ell ℓ samples along one image axis is ([7 Eq. (5.8] (Eqs. (8) to (10), (12) and (16); Goodman 2020))
ρ ( ℓ ) = ∣ ∫ − B / 2 B / 2 W ( f ) 2 e i 2 π f ℓ / f s d f ∣ ∫ − B / 2 B / 2 W ( f ) 2 d f , (11) \rho(\ell) = \frac{\left\lvert\int_{-B/2}^{B/2} W(f)^2 e^{i2\pi f\ell/f_s}\,\mathrm{d}f\right\rvert}{\int_{-B/2}^{B/2} W(f)^2\,\mathrm{d}f}, \tag*{(11)} ρ ( ℓ ) = ∫ − B /2 B /2 W ( f ) 2 d f ∫ − B /2 B /2 W ( f ) 2 e i 2 π f ℓ / f s d f , (11)
and the correlation at a two-dimensional lag is the product of the range and azimuth values. For example, the processing parameters of the TerraSAR-X stack of Section 5.1 (Table S11 of the supplement) predict a lag-one correlation of 0.695 in azimuth and 0.606 in range. We treat two pixels as sharing no speckle, and refer to them as lying beyond the speckle correlation length, if the predicted ∣ ρ q − p ∣ \lvert\rho_{q-p}\rvert ∣ ρ q − p ∣ is below ρ min = 0.025 \rho_{\min} = 0.025 ρ m i n = 0.025 , a bound tuned with the truncation step (Section 4.1).
Table S11. Processing parameters of the SLC data, with a raised-cosine (generalized Hamming) spectral weighting on both axes, and the lag-one speckle correlation that they predict through Eq. (11) , which the processing uses. The range sampling rate of Sentinel-1 is that of the data oversampled by a factor of two in range.
SPC under differential motion and subtracted model phases
SPC responds to signal differences as well as noise. For a noise-free relative phase that changes by δ \delta δ from one acquisition to the next,
S p q = ∣ sin ( N δ / 2 ) N sin ( δ / 2 ) ∣ 2 . (12) S_{pq} = \left|\frac{\sin(N\delta/2)}{N\sin(\delta/2)}\right|^2 . \tag*{(12)} S pq = N sin ( δ /2 ) sin ( N δ /2 ) 2 . (12)
Low SPC therefore does not imply noise, because two coherent pixels with different deformation can have low SPC.
A related case is important in practice. Suppose a model phase ψ p ( n ) \psi_p(n) ψ p ( n ) , such as a simulated topographic phase, is subtracted from every pixel before SPC is computed, and the model differs between neighbors by ψ p q ( n ) = ψ q ( n ) − ψ p ( n ) \psi_{pq}(n) = \psi_q(n) - \psi_p(n) ψ pq ( n ) = ψ q ( n ) − ψ p ( n ) . Under the null hypothesis of § 2.1.5 , the means in Eq. (9) become ξ n = g ( ∣ ρ ∣ ) e − i ψ p q ( n ) \xi_n = g(|\rho|)e^{-i\psi_{pq}(n)} ξ n = g ( ∣ ρ ∣ ) e − i ψ pq ( n ) , and
E [ Q p q ] = g ( ∣ ρ ∣ ) 2 Q [ e i ψ p q ] , (13) \mathbb{E}[Q_{pq}] = g(|\rho|)^2 Q\left[e^{i\psi_{pq}}\right], \tag*{(13)} E [ Q pq ] = g ( ∣ ρ ∣ ) 2 Q [ e i ψ pq ] , (13)
where Q [ e i ψ p q ] Q[e^{i\psi_{pq}}] Q [ e i ψ pq ] is the corrected SPC of the deterministic model difference alone. Where the model phase changes strongly between neighbors, as for simulated topographic phase in layover, the subtraction itself reduces SPC below the correlated-noise floor. § 3.1.2 shows this effect on real data.
Applications to masking and adaptive neighborhood sizing
This section describes the applications of SPC, and the steps they are combined with, in the order of the processing of a stack (Figure 2 ). Two per-pixel maps are computed first, before any SHP selection, the NSPC mask (§ 2.2.1 ) and the PCI (§ 2.2.2 ). For each center pixel, the search window is then scaled with the PCI (§ 2.2.3 ). Within it, the ensemble is collected by an SHP selection, either a pairwise one with thresholds that depend on the PCI (§ 2.2.4 and § 2.2.5 ) or the full window, which takes all eligible pixels of the window (§ 2.2.3 ) and serves as a reference for the pairwise ones. We count the full window among the SHP selections, although it tests no pixel. An adaptive truncation step , which can follow any SHP selection, may reduce the ensemble to a nested subset. In ANS it acts only where a gate on the PCI passes (§ 2.2.6 ). Together, the scaled window and the gated truncation step form the adaptive neighborhood sizing proposed here, ANS. The truncation step reduces the ensemble where a smaller nested subset agrees clearly better in phase with the center, and the scaled window gives a coherent center fewer candidates, since the number of looks needed for a given precision decreases with the coherence. The coherence matrix is then estimated from the returned ensemble in the same way for every SHP selection, with or without these steps (§ 2.2.8 ).
Figure 2. Processing of a stack. Per stack, the mask and the PCI are computed from fixed neighbors of each pixel, and the null quantile η 0 \eta_0 η 0 of the PCI and the threshold tables of FaSHPS and CMP from simulated clutter. Per center retained by the mask, the PCI scales the search window (S), sets the second-pass threshold of FaSHPS and the thresholds of CMP from these tables and gates the truncation step (T), which compares nested subsets of the ensemble with the phases of the core around the center. S and T together form ANS. The shaded steps are those introduced in this paper. Appendix § 11.1 gives the processing as an algorithm.
Figure 2 shows this processing with the three SHP selections compared here, the full window, FaSHPS of Jiang et al. [6 ] and CMP of Zhao et al. [24 ] , both described in § 2.2.5 . The mask step of Figure 2 applies to the real stacks, whereas the simulations apply no mask (§ 2.3.1 ). The simulations of § 2.3 mirror the real stacks in three configurations, named by sensor and number of acquisitions, TSX-44 for the TerraSAR-X stack with 44 acquisitions and S1-79 and S1-20 for the Sentinel-1 stack with 79 acquisitions and its substack of 20. Appendix § 11.1 gives the processing as an algorithm.
Neighborhood SPC mask
Let V \mathcal{V} V contain the pixels that have finite, nonzero observations in all acquisitions and that have not been excluded by any prior mask, and let N p = ( W p ∩ V ) ∖ { p } \mathcal{N}_p = (\mathcal{W}_p \cap\mathcal{V}) \setminus\{p\} N p = ( W p ∩ V ) ∖ { p } be the valid neighbors of p p p in a small neighborhood W p \mathcal{W}_p W p centered on p p p .
We define the neighborhood SPC as the mean of the square roots of the clipped corrected SPC,1
NSPC ( p ) = 1 ∣ N p ∣ ∑ q ∈ N p Q p q + (14) \operatorname{NSPC}(p) = \frac{1}{\lvert N_p \rvert} \sum_{q \in N_p} \sqrt{Q_{pq}^{+}} \tag*{(14)} NSPC ( p ) = ∣ N p ∣ 1 q ∈ N p ∑ Q pq + (14)
with NSPC ( p ) = 0 \operatorname{NSPC}(p) = 0 NSPC ( p ) = 0 for p ∉ V p \notin\mathcal{V} p ∈ / V or without valid neighbors. We use a 3 × 3 3 \times3 3 × 3 neighborhood throughout, so that all eight neighbors, the direct neighbors, are at lag one. A larger neighborhood would add neighbors at lags of two and more, where the SPC of zero-coherence clutter, which is correlated only through the point spread function, is much lower. In the TerraSAR-X stack of § 3.1 , the mean corrected SPC at lag two is a small fraction of that at lag one. NSPC would then lie much closer to zero in areas dominated by such clutter, which leaves little room for the values below the floor that the mask detects.
Value ranges Because the 3 × 3 3 \times3 3 × 3 neighborhood contains only lag-one neighbors, Eq. (10) sets a floor that zero-coherence clutter alone reaches. In the TerraSAR-X stack of § 3.1 , the NSPC values fall into three ranges. Their limits depend on the speckle correlation and on the number of acquisitions and are specific to that stack.
Above about 0.6. Agreement beyond shared speckle, i.e., coherent signal common to the pixel and its neighbors.
Around 0.43. The correlated-noise floor (§ 3.1.1 ).
Below about 0.3. Agreement below the floor. The simulations use a separable model of speckle correlation and temporal coherence, in which the covariance is the product of the temporal coherence and the speckle correlation. Under this model, decorrelation alone does not produce agreement below the floor, since the expected corrected SPC of direct neighbors is then at least g ( ∣ ρ ∣ ) 2 g(\lvert\rho\rvert)^2 g (∣ ρ ∣ ) 2 at any coherence
(Section S1.2 of the supplement). Within this model, it requires systematic differences between the phase histories of neighbors, namely differential motion, mixing of scatterers at different heights, or a subtracted model phase that the data does not share. Outside the separable model, noise that is less correlated than speckle, such as thermal noise, and the sidelobes of a bright scatterer, which reach adjacent samples with opposite signs (§ 3.1.2 ), also produce it.
Threshold and mask For a threshold h h h , the mask retains the pixels V h = { p ∈ V : NSPC ( p ) ≥ h } \mathcal{V}_h = \{p \in\mathcal{V} : \operatorname{NSPC}(p) \ge h\} V h = { p ∈ V : NSPC ( p ) ≥ h } . We set h h h by inspection of the NSPC histogram (§ 3.1.1 ), below the correlated-noise mode, at h = 0.25 h = 0.25 h = 0.25 for this stack. To check this choice, we simulated clutter with zero coherence between any two acquisitions, for the speckle correlation and the 44 acquisitions of the TerraSAR-X stack (TSX-44). In this clutter, NSPC has a median of 0.43, its 1 % and 0.1 % quantiles are 0.31 and 0.26, respectively, and it falls below 0.25 with a probability of 4 × 10 − 4 4 \times10^{-4} 4 × 1 0 − 4 .
Sensor dependence The floor depends on the speckle correlation, and hence on the sampling of the SLC data relative to its resolution. In simulated homogeneous clutter at zero coherence, the mean corrected SPC of neighbors at lag one along an image axis is lowest for range neighbors on TerraSAR-X, at 0.25, and highest for range neighbors on Sentinel-1, at 0.51, since that data is oversampled by a factor of two in range (§ 3.1 ). A fixed h h h therefore does not transfer between sensors, and we recommend setting h h h relative to the floor simulated for the measured or predicted ρ \rho ρ . The thresholds set here by inspection, h = 0.25 h = 0.25 h = 0.25 on TerraSAR-X and h = 0.30 h = 0.30 h = 0.30 on Sentinel-1, are 0.56 to 0.58 times the median NSPC of zero-coherence clutter simulated for the TerraSAR-X and Sentinel-1 stacks, and half this median can serve as a rough starting point. The prediction of ρ \rho ρ holds for SLCs on their original grid with known processing parameters, including a band-limited oversampling at its effective sampling rate. Coregistration, however, resamples every acquisition except the reference, which changes the correlation slightly (Section S4.1 of the supplement). The floor can also vary within a scene. Thermal noise is not shaped by the azimuth antenna pattern and is less correlated in azimuth than speckle, so pixels dominated by thermal noise have a lower floor than pixels dominated by zero-coherence clutter, a floor closer to the threshold. On Sentinel-1, a large part of the radar shadow falls below the threshold, with an NSPC lower still than that simulated for thermal noise (§ 3.1.2 ).
Scope of the mask The NSPC mask flags pixels that agree with their neighbors less than zero-coherence clutter with the speckle correlation does. The cause is either a systematic phase difference between neighbors, such as a subtracted topographic phase that does not match the phase of the data in layover, or noise that is less correlated than speckle, such as thermal noise in radar shadow. On both sensors the mask covers a large part of the geometric layover, and on Sentinel-1 also of the shadow (§ 3.1.2 ). It provides complementary information where the DEM is incomplete, needs no data beyond the stack after topographic phase removal, and is computed with eight inner products of length N N N per pixel, one per neighbor. However, isolated coherent targets and sharp deformation gradients can also have low NSPC (Eq. (12) ). Their retention should be checked when NSPC is used as a hard mask.
Treatment of masked pixels Where the mask precedes the SHP selection, as in the processing of the real stacks in this paper, masked pixels are not processed as centers, receive no estimate and enter no ensemble. Appendix § 11.1 gives their treatment in the PCI and in the truncation step.
Phase coherence indicator
The NSPC of a pixel averages SPC over its direct neighbors, which share speckle with the pixel. This gives NSPC the correlated-noise floor on which the mask relies, but it also makes NSPC respond weakly to low and moderate coherence. In simulated homogeneous clutter, the mean corrected SPC of adjacent pixels rises by at most 0.061 from zero coherence to a coherence of 0.5 between consecutive acquisitions, on both sensors and in both directions. An indicator of local coherence at a pixel therefore uses neighbors that share no speckle with it. Let
η ~ p = 1 ∣ R p ∣ ∑ q ∈ R p Q p q + , (15) \tilde{\eta}_p = \frac{1}{|\mathcal{R}_p|}\sum_{q\in\mathcal{R}_p}\sqrt{Q^+_{pq}}, \tag*{(15)} η ~ p = ∣ R p ∣ 1 q ∈ R p ∑ Q pq + , (15)
where the ring R p \mathcal{R}_p R p contains the 24 pixels nearest to p p p beyond the speckle correlation length, i.e., as many pixels as a 5 × 5 5 \times5 5 × 5 window holds besides its center, though none of them shares speckle with p p p . The region beyond the correlation length is F p = { q : ∣ ρ q − p ∣ < ρ min } \mathcal{F}_p = \{q : |\rho_{q-p}| < \rho_{\min}\} F p = { q : ∣ ρ q − p ∣ < ρ m i n } with the predicted correlation ρ \rho ρ , a set that excludes p p p itself. Nearness is measured by the normalized distance of Eq. (16) (§ 2.2.3 ), and the order it gives the offsets depends only on the ratio of the semiaxes b 1 b_1 b 1 and b 2 b_2 b 2 of the search window. Offsets tied in this distance at the cut are taken in point-symmetric pairs, so that the ring is symmetric about p p p . In the experiments of this paper, the ring pixels lie at 3 to 4.1 pixels from p p p on TerraSAR-X and, mostly along range, at 4 to 12 pixels on Sentinel-1, where the data is oversampled in range and the search window is elongated in range. The phase coherence indicator (PCI) η p \eta_p η p is the mean of η ~ q \tilde{\eta}_q η ~ q over the 25 pixels q q q of the 5 × 5 5 \times5 5 × 5 window centered on p p p . The point coherence estimation of Vecchioli et al. (2023) also rests on pairs of points with independent phase noise (§ 2.1.2 ) and solves for the coherence of every point, whereas the PCI averages over a fixed ring and a 5 × 5 5 \times5 5 × 5 neighborhood and thus describes the coherence of p p p together with that of its surroundings. These surroundings span an area of 13 × 13 13 \times13 13 × 13 pixels (about 26 × 25 26 \times25 26 × 25 m in azimuth and ground range) around p p p on TerraSAR-X and 7 × 29 7 \times29 7 × 29 pixels (about 98 × 59 98 \times59 98 × 59 m) on Sentinel-1. In the processing of this paper (Algorithm A.1), η \eta η is computed once for every pixel of the image, and all later uses of the PCI, including Table 1 , refer to this value.
Table 1. PCI η \eta η (§ 2.2.2 ) at the simulated distributed centers (evaluation scenes, § 2.3.1 ) by the coherence c c c between consecutive acquisitions, for the configurations TSX-44, S1-79 and S1-20 (44 TerraSAR-X, 79 and 20 Sentinel-1 acquisitions, Table 2 ): weighted median, g g g of Eq. (7) , the mean cosine of the single-look phase error, at the long-term coherence c 2 c^2 c 2 of Eq. (27) , and the null quantile η 0 \eta_0 η 0 , the 99th percentile of η \eta η in simulated homogeneous clutter with zero coherence (calibration scenes).
To relate η \eta η to the coherence, consider neighbors with the same coherences ∣ γ s t ∣ |\gamma_{st}| ∣ γ s t ∣ and the same true phase history, for which Eq. (8) gives E [ Q p q ] \mathbb{E}[Q_{pq}] E [ Q pq ] as the mean of g ( ∣ γ s t ∣ ) 2 g(|\gamma_{st}|)^2 g ( ∣ γ s t ∣ ) 2 over the pairs. Since η \eta η averages the roots Q p q + \sqrt{Q^+_{pq}} Q pq + instead of taking the root of the mean of the unclipped Q p q Q_{pq} Q pq , it differs from E [ Q p q ] \sqrt{\mathbb{E}[Q_{pq}]} E [ Q pq ] . In the simulations (§ 2.3.1 ), the coherence between consecutive acquisitions is c c c , and it decays toward c 2 c^2 c 2 as the interval grows (Eq. (27) ). In TSX-44 at c = 0.5 c = 0.5 c = 0.5 , E [ Q p q ] \sqrt{\mathbb{E}[Q_{pq}]} E [ Q pq ] is 0.28, and the weighted median of η \eta η is 0.22. Table 1 gives the weighted median of η \eta η at the distributed centers of the simulations (§ 2.3.1 ), together with g ( c 2 ) g(c^2) g ( c 2 ) , the value of g g g in Eq. (7) at the long-term coherence. The median rises monotonically with the coherence between consecutive acquisitions, its value at zero coherence decreases with the number of acquisitions, and from c = 0.5 c = 0.5 c = 0.5 on it lies within 0.06 of g ( c 2 ) g(c^2) g ( c 2 ) . The root E [ Q p q ] \sqrt{\mathbb{E}[Q_{pq}]} E [ Q pq ] lies above g ( c 2 ) g(c^2) g ( c 2 ) , since the shorter intervals have higher coherences, and the median of η \eta η lies below the root at every c > 0 c > 0 c > 0 . At a given coherence, however, η \eta η is lower in dense fringes, since the phase difference between the pixel and its ring neighbors then changes over the stack (Eq. (8) ). In this paper, η \eta η sets the thresholds of the pairwise SHP selections (§ 2.2.5 ), scales the search window (§ 2.2.3 ) and gates the truncation step (§ 2.2.6 ), and it is shown in the maps of § 3 .
PCI-scaled search window
For an eligible center p p p , let B p \mathcal{B}_p B p contain the eligible pixels in a bounded search window around p p p , including p p p . Eligible pixels belong to V h \mathcal{V}_h V h if masking is applied, and to V \mathcal{V} V otherwise. The search window has the same size at every center of a stack, and B p \mathcal{B}_p B p holds fewer pixels only where the window extends beyond the image or contains pixels that are not eligible. The user chooses its size large enough for the decorrelated parts of a scene, which need many looks. A coherent center needs fewer looks for the same precision, and the distant members of an SHP selection in the unscaled search window are those most likely to carry a different phase. The truncation step (§ 2.2.6 ) reduces such an ensemble only where a smaller nested subset agrees clearly better with the center, and the scaled window complements it by limiting the number of candidates at a coherent center before the SHP selection. Since the PCI (§ 2.2.2 ) rises with the coherence of the center (Table 1 ), it is used here to shrink the search window of each center.
The window is shrunk only above the null quantile η 0 \eta_0 η 0 of the PCI, the 99th percentile of η \eta η in simulated homogeneous clutter with zero coherence, computed for each configuration on the calibration scenes (§ 2.3.1 ). The null quantile is 0.066 for TSX-44, and 0.050 and 0.100 for S1-79 and S1-20, respectively.
(Table 1 ). For the point spread functions of TerraSAR-X and Sentinel-1, η 0 N \eta_0\sqrt{N} η 0 N is 0.44 to 0.45 in all three configurations, i.e., η 0 \eta_0 η 0 is nearly proportional to 1 / N 1/\sqrt{N} 1/ N .
For a pixel q q q , let d q = q − p d_q=q-p d q = q − p be its offset from the center, with the components d q , 1 d_{q,1} d q , 1 in azimuth and d q , 2 d_{q,2} d q , 2 in range, in pixels. Its normalized distance from the center is given by
∥ d q ∥ b 2 = ( d q , 1 b 1 ) 2 + ( d q , 2 b 2 ) 2 , (16) \lVert d_q\rVert_b^2=\left(\frac{d_{q,1}}{b_1}\right)^2+\left(\frac{d_{q,2}}{b_2}\right)^2, \tag*{(16)} ∥ d q ∥ b 2 = ( b 1 d q , 1 ) 2 + ( b 2 d q , 2 ) 2 , (16)
where b 1 b_1 b 1 and b 2 b_2 b 2 are the semiaxes of the search window in azimuth and range, i.e., half the side lengths of the rectangle in Table 2 , so that B p B_p B p contains the eligible pixels with ∥ d q ∥ b ≤ 1 \lVert d_q\rVert_b\leq1 ∥ d q ∥ b ≤ 1 . With the null quantile η 0 \eta_0 η 0 of the PCI, the relative PCI and the scale of the window are
Table 2. Configurations and settings, identical in simulation and on real data except that the simulations apply no mask. The search window is the ellipse inscribed in the stated rectangle (azimuth × range). The window of TSX-44 (44 TerraSAR-X acquisitions) is larger because that stack decorrelates more strongly and therefore needs more looks. The speckle correlation is the lag-one value that the processing uses (Table S11 of the supplement), from the prediction from the processing parameters of the sensor (Eq. (11) , § 2.1.5 and § 3.1.1 ). The core, the guard set and the ring of the PCI follow from this prediction at every lag and from ρ min \rho_{\min} ρ m i n . One fringe per window diameter is given as the difference of line-of-sight displacement across the window and as its rate over the time span of the corresponding real stack (5.24, 2.96 and 0.79 years for TSX-44, S1-79 and S1-20, respectively, § 3 ). The null quantile and the thresholds were calibrated in homogeneous clutter of the calibration scenes, κ \kappa κ , λ \lambda λ and ρ min \rho_{\min} ρ m i n were tuned on the tuning scenes, and the fixed values were set without tuning and changed only in the checks of Sections S3.1 and § 14.1 of the supplement. The mask threshold was set by inspection of the NSPC histogram of each real stack (§ 2.2.1 ). Section S3.1 of the supplement discusses the fixed values. An equals sign repeats the value of TSX-44.
η p ′ = min ( 1 , max ( 0 , η p − η 0 1 − η 0 ) ) , τ p = 1 − η p ′ ( 1 − τ min ) , (17) \eta'_p=\min\left(1,\max\left(0,\frac{\eta_p-\eta_0}{1-\eta_0}\right)\right),\qquad\tau_p=1-\eta'_p(1-\tau_{\min}), \tag*{(17)} η p ′ = min ( 1 , max ( 0 , 1 − η 0 η p − η 0 ) ) , τ p = 1 − η p ′ ( 1 − τ m i n ) , (17)
respectively, where τ min \tau_{\min} τ m i n is the normalized distance of the n min n_{\min} n m i n -th nearest pixel. The PCI-scaled window B p ′ B'_p B p ′ contains the pixels of B p B_p B p with ∥ d q ∥ b ≤ τ p \lVert d_q\rVert_b\leq\tau_p ∥ d q ∥ b ≤ τ p , and at least the n min n_{\min} n m i n nearest ones. Its semiaxes thus decrease linearly with the PCI, from those of B p B_p B p at η p ≤ η 0 \eta_p\leq\eta_0 η p ≤ η 0 to those of the n min n_{\min} n m i n nearest pixels at η p = 1 \eta_p=1 η p = 1 , so that the number of its pixels falls about quadratically.
The value n min = 100 n_{\min}=100 n m i n = 100 is fixed without tuning. Since no center reaches the window of η p = 1 \eta_p=1 η p = 1 , in the simulations or on the real stacks, n min n_{\min} n m i n acts only through the slope of τ p \tau_p τ p in η p \eta_p η p . With n min = 50 n_{\min}=50 n m i n = 50 or 200 the error of the full window and FaSHPS with ANS changes by at most 0.4% over all centers (TSX-44 and S1-79, § 14.1 of the supplement).
An SHP selection then works within B p ′ B'_p B p ′ in place of B p B_p B p .
Amplitude tests under temporal correlation
Amplitude tests for statistically homogeneous pixels treat the N N N amplitudes of a pixel as independent samples. Distribution tests compare the amplitudes of two pixels as samples of one distribution (Ferretti et al., [3 ] ; Parizzi and Brcic, [14 ] ), and the test of Jiang et al. [6 ] Eq. (7)) compares their temporal means with a confidence interval derived for $N$ independent samples, which they scale with one coefficient of variation per scene ({xref}Section 3.5`). For a distributed scatterer, independence holds only if the speckle decorrelates between acquisitions. Otherwise, the amplitude of a pixel changes little from one acquisition to the next. Two pixels of the same homogeneous surface then have different, nearly constant amplitudes, and a test that assumes independent samples wrongly rejects them. Under the CCG model with coherence matrix Γ \Gamma Γ and equal mean intensity in all acquisitions, the intensities of one pixel in acquisitions s s s and t t t have correlation coefficient ∣ γ s t ∣ 2 |\gamma_{st}|^2 ∣ γ s t ∣ 2 (Goodman, 2020, p. 53, Eq. (3.97)). For the mean of correlated observations, equating its variance to that of the mean of independent samples defines the effective sample size (Thiébaux and Zwiers, 1984, pp. 800–801, Eq. (2.1)). For the temporal mean intensity it is
N e f f = N 2 ∑ s , t ∣ γ s t ∣ 2 , (18) N_{\mathrm{eff}} = \frac{N^2}{\sum_{s,t} |\gamma_{st}|^2}, \tag*{(18)} N eff = ∑ s , t ∣ γ s t ∣ 2 N 2 , (18)
where the sum includes the terms s = t s=t s = t with ∣ γ s s ∣ = 1 |\gamma_{ss}|=1 ∣ γ ss ∣ = 1 . Goodman (2020) gives the continuous counterpart of Eq. (18) for temporally integrated speckle. For the temporal mean amplitude, which the tests below use, Eq. (18) is a close approximation. We have N e f f = N N_{\mathrm{eff}}=N N eff = N for a fully decorrelated target and N e f f → 1 N_{\mathrm{eff}}\to1 N eff → 1 for a fully coherent one. A test calibrated for N N N samples attains its significance level only in the first case. The more coherent the target, the more often such a test rejects homogeneous neighbors, which is contrary to the desired behavior, because the homogeneity of a target does not depend on its coherence. The thresholds of FaSHPS, and those of CMP, whose statistic also contains the ratio of the temporal mean intensities, are therefore set in temporally correlated clutter, and the second-pass threshold of FaSHPS and the thresholds of CMP depend on the PCI of the center, which rises with its coherence (§ 2.2.5 ).
Pairwise SHP selection and connectivity
Besides the full window, this paper combines ANS with two pairwise SHP selections. FaSHPS compares each candidate q q q in the search window with p p p on the temporal mean amplitudes, CMP on the temporal mean intensities and the single-look phase differences, and we call these two tests the SHP tests. Phase-based pixel selection has also been formulated as tests on covariance matrices (Dong et al., 2018; Song et al., 2017) and, without reference to the center, as a clustering by the trigonometric moments of residual phases (Yao and Balz, 2024), which groups pixels by their decorrelation rather than by the agreement of their phases with those of the center. These methods are not compared here.
Amplitude test We use the statistic of Jiang et al. [6 ] as the representative of the amplitude tests of SqueeSAR-type frameworks [3 ] . It shares their assumption of independent samples (§ 2.2.4 ) and requires a few operations per candidate, independent of N N N (§ 2.2.9 ). Its authors report improvements over previous methods, among them the Kolmogorov–Smirnov SHP selection of SqueeSAR, in edge preservation, retention of resolution, noise reduction, bias mitigation and computational efficiency (their Section VII). Following their Sections III-A and IV, let A ˉ q \bar{A}_q A ˉ q be the temporal mean amplitude of pixel q q q after the removal of outlying amplitudes with a revised boxplot, and
T p q = ∣ A ˉ q − μ p ∣ c A μ p / N , c A = 0.52 , (19) T_{pq} = \frac{\left|\bar{A}_q-\mu_p\right|}{c_A\mu_p/\sqrt{N}}, \qquad c_A=0.52, \tag*{(19)} T pq = c A μ p / N A ˉ q − μ p , c A = 0.52 , (19)
where c A c_A c A is the coefficient of variation of a Rayleigh amplitude, so that T p q ≤ z T_{pq}\le z T pq ≤ z is the confidence interval of their (7) .
The test has two passes, of which the first only gives the second a reference mean from more pixels than the center alone. In the first pass, μ p = A ˉ p \mu_p=\bar{A}_p μ p = A ˉ p , and the pixels with T p q ≤ z 1 T_{pq}\le z_1 T pq ≤ z 1 form, with the center, the reference set of p p p . In the second pass, μ p \mu_p μ p is the mean of A ˉ q \bar{A}_q A ˉ q over the reference set, every candidate of the window is tested against it, and a candidate passes if T p q ≤ z α T_{pq}\le z_\alpha T pq ≤ z α , where α \alpha α is the significance level of the test. The pixels that pass form the ensemble O p \mathcal{O}_p O p through their eight-connected component that contains p p p , whereas the original test accepts any pixel of the window that passes.
The original thresholds z 1 z_1 z 1 and z α z_\alpha z α are normal quantiles of a 50 % 50\% 50% and 1 − α 1-\alpha 1 − α interval, respectively, for N N N independent samples, with α = 5 % \alpha=5\% α = 5% in their test, and both are scaled with a coefficient of variation of the temporal mean estimated once per scene (their Section III-B). We use them only for comparison, at the significance level of 1 % 1\% 1% used in this paper (Section S2.2 of the supplement).
By § 2.2.4 , a threshold for N N N independent samples attains its significance level only at zero coherence, and a threshold set at one coherence loses it at another. A single value per scene cannot hold the level at every center of a scene of mixed coherence. The outlier removal and, on oversampled data, the shared speckle of adjacent pixels change the distribution of T p q T_{pq} T pq as well. We therefore set both thresholds by simulating homogeneous clutter with the stack size, the search window and the processing parameters of the SLC data (§ 2.1.5 ), once per stack and before the centers are processed. The first-pass threshold z 1 z_1 z 1 admits half of the homogeneous candidates to the reference set at a coherence of 0.5, in the middle of the simulated coherences, more at lower and fewer at higher coherence. It is the same at every center, since the significance level and its dependence on the coherence lie in the second pass. For the second pass, the 1 − α 1-\alpha 1 − α quantile of T p q T_{pq} T pq is tabulated against the median PCI of the clutter at eight coherences from 0 to 0.95. Each center takes the threshold z α ( η p ) z_{\alpha}(\eta_p) z α ( η p ) interpolated linearly at its own PCI (§ 2.2.2 ), which rises with the coherence (Table 1 ). The threshold table contains no real data, which enters only through the PCI of the center (Section S2.2 of the supplement). The statistic keeps the theoretical c A c_A c A instead of a value estimated from the scene. With simulated thresholds, c A c_A c A only fixes the scale of T p q T_{pq} T pq , so that the test is the original one with other thresholds. At a significance level of α = 1 % \alpha= 1\% α = 1% , the test then rejects 0.6 to 1.8% of the homogeneous candidates at every simulated coherence, including coherences between those of the calibration. The original thresholds reject 1.9 to 27%, most at c = 0.95 c = 0.95 c = 0.95 with 10 to 27% (three configurations, Figure 3 and Table S1 of the supplement). We keep the name FaSHPS for this adjusted SHP selection.
Figure 3. With the threshold that depends on the PCI, FaSHPS rejects homogeneous candidates at close to the significance level at every coherence, and with the original thresholds more often, most at the highest coherence. Share of the homogeneous candidates outside the guard set that FaSHPS rejects with the original thresholds (dashed) and with the threshold that depends on the PCI η \eta η (solid), at a significance level of 1 % (dotted line), in simulated homogeneous clutter by the coherence c c c between consecutive acquisitions (Eq. (27) ). The guard set holds the pixels that share speckle with the center or with a pixel that does (§ 2.2.6 ). The shares are means over scenes generated from the seeds of the evaluation scenes, for each of the configurations TSX-44, S1-79 and S1-20 (44 TerraSAR-X, 79 and 20 Sentinel-1 acquisitions, § 2.3.1 ). The original thresholds are those of Jiang et al. [6 ] at this significance level instead of their 5 %, with a coefficient of variation estimated per scene, which in these scenes of a single coherence fits every center (Section S2.2 of the supplement). Table S1 of the supplement gives the values, also for CMP and at coherence steps.
Table S1. Share in percent of the candidates of the same class outside the guard set (§ 2.2.6 ) that each test rejects at a significance level of 1 %, in homogeneous clutter and at coherence steps, mean over scenes generated from the seeds of the evaluation scenes, for FaSHPS and CMP with the threshold that depends on the PCI of the center and with the original thresholds (orig.). A coherence step joins two classes of equal amplitude and phase but different coherence, and the candidates of the same class are those in the class of the center. The thresholds that depend on the PCI are tabulated at the PCI of the calibration coherences 0, 0.1, 0.2, 0.35, 0.5, 0.65, 0.8 and 0.95 and interpolated linearly between them (Section S2.2).
Test of CMP CMP ([24 ] ) compares two pixels by a symmetrized Kullback–Leibler divergence over the single-reference interferograms, which combines the ratio of their temporal mean intensities with a weighted mean cosine of their single-look phase differences, with weights from an assumed coherence of each interferogram. Of the terms in the statistics of the SHP tests, this phase term is the one most closely related to SPC, but it requires a choice of interferograms, here the single-reference ones, and a coherence model, whereas SPC requires neither. As in FaSHPS, the test has two passes, the second against the mean intensity and the multilooked phases of the reference set formed by the first, and the ensemble is the eight-connected component of the center ([24 Section 2.2.3] ). Both of its thresholds are tabulated over the PCI from the same simulated clutter as the second-pass threshold of FaSHPS, at the same significance level, with the decay constant of its coherence model fitted to the stack (Section S2.1 of the supplement). That section also gives the statistic and the coherence model, of which we leave out the term of the perpendicular baseline.
Limitation at high coherence At high coherence, the speckle of a pixel changes little between acquisitions, and its temporal mean amplitude scatters about the mean amplitude of its surface nearly as much as a single amplitude. A test on the mean amplitudes must then either accept candidates over a wide range of brightness or reject homogeneous candidates more often than its significance level, and this applies to any test that uses only the amplitudes. The phase term of CMP on its own rejects practically no candidate (Section S4.8 of the supplement), so that candidates of similar brightness but different phase are left to the truncation step.
Minimum size If O p \mathcal{O}_p O p has fewer than L min = 25 L_{\min} = 25 L m i n = 25 members, the center is insufficient, and its processing stops. The center keeps its single-look data. In the comparisons below, the error of such a center is computed from the single-look matrix u p u p H \mathbf{u}_p\mathbf{u}_p^{\mathrm{H}} u p u p H , so that all errors refer to the same pixels, and the share of insufficient centers is reported for the SHP selections with and without ANS (§ 13.3.11 of the supplement).
Truncation step
An SHP selection returns every compatible pixel it finds, which is the desired behavior in a homogeneous area, where all members share the true phase history of the center, because there the estimate improves with the number of looks at any coherence. Where coherence is high and the phase varies rapidly in space, however, members far from the center carry a different true phase and bias the estimate, and where the center is dominated by a single stable scatterer, its own phase is more accurate than an average over hundreds of pixels with other heights and histories (§ 2.2.7 ). A single value of Q p q Q_{pq} Q pq cannot distinguish these situations from decorrelation, because a decorrelated neighbor and a coherent neighbor on a different fringe both have low SPC (Eqs. (8) and (12) ). The spatial pattern of Q p q Q_{pq} Q pq around the center can. Under decorrelation, Q p q Q_{pq} Q pq is low at all offsets beyond the speckle correlation length, and under dense fringes it is high near the center and decays with distance, fastest along the phase gradient.
The adaptive truncation step takes an ensemble selected beforehand and returns either the ensemble itself or a subset of it. It orders the members using the SPC of the center with the surrounding pixels, scores nested subsets of increasing size up to the whole ensemble by how well they describe the center, and then decides which size to return. In a homogeneous area, every added member reduces the noise of the multilook phases, so that the score rises with the size up to estimation noise, and the ensemble is truncated only where this noise happens to favor a smaller subset. In dense fringes, the score attains its maximum at a small size and then decreases, and the ensemble is truncated. The three stages (ordering, score and decision) are described below. The step can follow any SHP selection, and in § 2.3 it is applied to the ensembles of the full window, FaSHPS and CMP.
Ordering of the members Members are ordered so that those most likely to share the true phase of the center come first. The simplest such order, the distance order , sorts them by the normalized distance from the center (Eq. (16) ). The corrected SPC Q p q Q_{pq} Q pq of the center with the pixels q q q of F p \mathcal{F}_p F p within the search window (F p \mathcal{F}_p F p of § 2.2.2 , B p ′ B'_p B p ′ of § 2.2.3 with the PCI-scaled window) suggests an alternative. It stays high along a fringe and falls quickly across it. With the offsets d q d_q d q of Eq. (16) , the second-moment matrix of the offsets, weighted with these values, is
M p = ∑ q ∈ F p ∩ B p ′ Q p q d q d q T ∑ q ∈ F p ∩ B p ′ Q p q . (20) M_p = \frac{\sum_{q \in\mathcal{F}_p \cap\mathcal{B}'_p} Q_{pq} d_q d_q^{\mathsf{T}}}{\sum_{q \in\mathcal{F}_p \cap\mathcal{B}'_p} Q_{pq}}. \tag*{(20)} M p = ∑ q ∈ F p ∩ B p ′ Q pq ∑ q ∈ F p ∩ B p ′ Q pq d q d q T . (20)
Like a covariance matrix of the offsets, M p M_p M p describes an ellipse, long in the direction in which the center retains a high SPC with distant pixels and short in the direction in which the SPC decreases quickly. Sorting the members by d q T M p − 1 d q d_q^{\mathsf{T}} M_p^{-1} d_q d q T M p − 1 d q , the squared Mahalanobis distance (cf. Akyilmaz and Leloglu, 2017, Eq. (2) ), places pixels along a fringe before pixels at the same Euclidean distance across it. We call this the ellipse order . Where M p M_p M p is not positive definite or the SPC of the center carries too little weight, the distance order is used (Section S2.3 of the supplement). The ellipse order is fixed by design, and Section S3.1 of the supplement compares it with the distance order, which has a higher error in dense fringes. In either case, ν ( q ) \nu(q) ν ( q ) denotes the position of q q q in the sorted ensemble, counted from one.
Score of a candidate ensemble The score assesses nested subsets of the ensemble O p \mathcal{O}_p O p by how well their multilooked phases agree with those of the core, the pixels that share their speckle with the center. Here O p \mathcal{O}_p O p denotes the ensemble passed to the truncation step, i.e., the ensemble of any SHP selection, sorted as described above.
The core is C p = { q : ∣ ρ q − p ∣ ≥ ρ min } \mathcal{C}_p = \{q : |\rho_{q-p}| \ge\rho_{\min}\} C p = { q : ∣ ρ q − p ∣ ≥ ρ m i n } , the pixels within the speckle correlation length of p p p , including p p p , i.e., the complement of F p \mathcal{F}_p F p (§ 2.2.2 ). This gives a core of 31 pixels on TerraSAR-X and of 55 pixels on Sentinel-1 (Table 2 ). Taking the core instead of p p p alone lowers the noise of the score, on the assumption that the true phase histories of the core pixels, which are not known, are similar to that of p p p .
Pixels whose speckle is correlated with that of a core pixel agree with the core whatever their true phase. If they were included in the multilook, small subsets would agree with the core because they contain the core and the pixels whose speckle is correlated with it, and the ensemble would be truncated in zero-coherence clutter. These pixels form the guard set G p = { q : ∣ ρ q − p ′ ∣ ≥ ρ min for some p ′ ∈ C p } ⊃ C p \mathcal{G}_p = \{q : |\rho_{q-p'}| \ge\rho_{\min}\ \text{for some } p' \in\mathcal{C}_p\} \supset\mathcal{C}_p G p = { q : ∣ ρ q − p ′ ∣ ≥ ρ m i n for some p ′ ∈ C p } ⊃ C p , and they are excluded from the score.
Let q 1 , q 2 , … q_1, q_2, \ldots q 1 , q 2 , … be the pixels of O p ∖ G p \mathcal{O}_p \setminus\mathcal{G}_p O p ∖ G p in the sorted order. The score is evaluated for the nested subsets
E k = { q 1 , … , q L k } , k = 1 , … , K , (21) \mathcal{E}_k = \{q_1, \ldots, q_{L_k}\}, \qquad k = 1, \ldots, K, \tag*{(21)} E k = { q 1 , … , q L k } , k = 1 , … , K , (21)
with sizes L k = ⌊ L 1 1.5 k − 1 ⌋ L_k = \left\lfloor L_1 1.5^{k-1} \right\rfloor L k = ⌊ L 1 1. 5 k − 1 ⌋ from L 1 = L min = 25 L_1 = L_{\min} = 25 L 1 = L m i n = 25 and, as the last size, L K = ∣ O p ∖ G p ∣ L_K = |\mathcal{O}_p \setminus\mathcal{G}_p| L K = ∣ O p ∖ G p ∣ . Each size L k L_k L k with k < K k < K k < K lies at least a factor 1.5 below $L_K (Section S2.3 of the supplement), so that with fewer than ⌈ 1.5 L min ⌉ = 38 \lceil1.5 L_{\min} \rceil= 38 ⌈ 1.5 L m i n ⌉ = 38 members outside the guard set there is no smaller candidate and the truncation step is skipped. To form the multilooked phases of these subsets, let A \mathcal{A} A be a set of R = 8 R = 8 R = 8 reference acquisitions spread evenly over the stack, and let I \mathcal{I} I be the set of unordered pairs of distinct acquisitions of which at least one belongs to A \mathcal{A} A . It has ∣ I ∣ = R ( N − 1 ) − R ( R − 1 ) / 2 |\mathcal{I}| = R(N - 1) - R(R - 1)/2 ∣ I ∣ = R ( N − 1 ) − R ( R − 1 ) /2 elements, so that the cost of the multilooks grows with R N RN R N instead of N 2 / 2 N^2/2 N 2 /2 . With a single reference acquisition, the outcome of the truncation step would change with the choice of that acquisition, and R = 8 R = 8 R = 8 is a trade-off between this cost and the phase linking error (Sections S4.9 and § 14.1 of the supplement). We write each pair as ( n , r ) (n,r) ( n , r ) with r ∈ A r \in\mathcal{A} r ∈ A . For ( n , r ) ∈ I (n,r) \in\mathcal{I} ( n , r ) ∈ I , the multilooked interferometric phasor of acquisitions n n n and r r r over E k \mathcal{E}_k E k and the multilooked interferogram of the core are
e ^ k ( n , r ) = ∑ q ∈ E k u q ( n ) u q ( r ) ‾ ∣ ∑ q ∈ E k u q ( n ) u q ( r ) ‾ ∣ , v ‾ c ( n , r ) = 1 ∣ C p ∣ ∑ p ′ ∈ C p u p ′ ( n ) u p ′ ( r ) ‾ , (22) \widehat{e}_k(n,r) = \frac{\sum_{q \in\mathcal{E}_k} u_q(n)\overline{u_q(r)}}{\left|\sum_{q \in\mathcal{E}_k} u_q(n)\overline{u_q(r)}\right|}, \qquad\overline{v}_c(n,r) = \frac{1}{|\mathcal{C}_p|}\sum_{p' \in\mathcal{C}_p} u_{p'}(n)\overline{u_{p'}(r)}, \tag*{(22)} e k ( n , r ) = ∑ q ∈ E k u q ( n ) u q ( r ) ∑ q ∈ E k u q ( n ) u q ( r ) , v c ( n , r ) = ∣ C p ∣ 1 p ′ ∈ C p ∑ u p ′ ( n ) u p ′ ( r ) , (22)
respectively. The score is the agreement of the two in the pair form of corrected SPC (Eq. (5) ), over the pairs of I \mathcal{I} I instead of all pairs, as the mean of the terms y k ( n , r ) y_k(n,r) y k ( n , r ) ,
a k = 1 ∣ I ∣ ∑ ( n , r ) ∈ I y k ( n , r ) , y k ( n , r ) = Re [ v ‾ c ( n , r ) e ^ k ( n , r ) ] , (23) a_k = \frac{1}{|\mathcal{I}|}\sum_{(n,r) \in\mathcal{I}} y_k(n,r), \qquad y_k(n,r) = \operatorname{Re}\left[\overline{v}_c(n,r)\widehat{e}_k(n,r)\right], \tag*{(23)} a k = ∣ I ∣ 1 ( n , r ) ∈ I ∑ y k ( n , r ) , y k ( n , r ) = Re [ v c ( n , r ) e k ( n , r ) ] , (23)
i.e., the mean over the pairs of the cosine of the difference between the multilooked phases of E k \mathcal{E}_k E k and of the core, each multiplied by ∣ v ˉ C ( n , r ) ∣ |\bar{v}_{C}(n,r)| ∣ v ˉ C ( n , r ) ∣ , the phase-only coherence of the core in that pair (Eq. (2) with W = C p \mathcal{W} = C_p W = C p ). Pairs in which the core has decorrelated thus weigh less.
For a single core pixel p ′ p' p ′ , a one-pixel ensemble E k = { q } \mathcal{E}_k = \{q\} E k = { q } and all acquisitions in A \mathcal{A} A , Eq. (5) gives a k = Q p ′ q a_k = Q_{p'q} a k = Q p ′ q , the mean cosine agreement of the interferometric phases of p ′ p' p ′ and q q q . The score generalizes this to the multilooked interferograms of E k \mathcal{E}_k E k and of the core, which need not be consistent with a single phase history, and to the pairs of I \mathcal{I} I . Like Q p q Q_{pq} Q pq , the score has zero mean when the phases of the core are unrelated to those of the ensemble, so, unlike S p q S_{pq} S pq , it needs no bias correction. Section S5 of the supplement gives its expectation over the speckle of the core and shows that, in the simulations, SHP selections and orders that use the phases of the center do not bias it.
Decision. The decision compares the gain of the highest score (Eq. (23) ) over the score of the whole ensemble with the noise of that gain, and the highest score with its own noise. Let a max = max k a k a_{\max} = \max_k a_k a m a x = max k a k , attained at k = k max k = k_{\max} k = k m a x . The truncation step retains the whole ensemble unless a smaller subset scores higher than the whole ensemble by more than the margin ζ \zeta ζ and its score exceeds λ \lambda λ times its noise σ k max \sigma_{k_{\max}} σ k m a x (Eq. (25) below):
E ⋆ = { O p if a max − a K ≤ ζ or a max ≤ λ σ k max , { q ∈ O p : ν ( q ) ≤ ν ( q L k ⋆ ) } otherwise , (24) \mathcal{E}^{\star} =
\begin{cases}
\mathcal{O}_p & \text{if } a_{\max} - a_K \le\zeta\text{ or } a_{\max} \le\lambda\sigma_{k_{\max}}, \\
\{q \in\mathcal{O}_p : \nu(q) \le\nu(q_{L_{k^\star}})\} & \text{otherwise},
\end{cases} \tag*{(24)} E ⋆ = { O p { q ∈ O p : ν ( q ) ≤ ν ( q L k ⋆ )} if a m a x − a K ≤ ζ or a m a x ≤ λ σ k m a x , otherwise , (24)
with k ⋆ = max { k : a k ≥ a max − ζ / 3 } k^\star= \max\{k : a_k \ge a_{\max} - \zeta/3\} k ⋆ = max { k : a k ≥ a m a x − ζ /3 } , the largest size whose score lies within the tolerance ζ / 3 \zeta/3 ζ /3 of the maximum. A truncated ensemble keeps the guard pixels that come before q L k ⋆ q_{L_{k^\star}} q L k ⋆ and at least L 1 L_1 L 1 members outside the guard set, so that it never falls below the minimum size L min L_{\min} L m i n of § 2.2.5 .
The margin is proportional to the noise of the gain Δ a = a k max − a K \Delta a = a_{k_{\max}} - a_K Δ a = a k m a x − a K , the mean of the term differences Δ y ( n , r ) = y k max ( n , r ) − y K ( n , r ) \Delta y(n,r) = y_{k_{\max}}(n,r) - y_K(n,r) Δ y ( n , r ) = y k m a x ( n , r ) − y K ( n , r ) over I \mathcal{I} I . Where the core carries information about the true phase, terms that share an acquisition are correlated, so that the standard deviation σ Δ a \sigma_{\Delta a} σ Δ a of Δ a \Delta a Δ a is estimated by a jackknife over acquisitions (Efron and Stein [1 ] ), given in Eq. (39) of the supplement. The margin is ζ = κ σ Δ a \zeta= \kappa\sigma_{\Delta a} ζ = κ σ Δ a , with the margin factor κ = 0.5 \kappa= 0.5 κ = 0.5 . The decision does not explicitly weigh the looks that truncation gives up. Instead, the margin lets a smaller ensemble trigger truncation only if its gain in agreement with the core exceeds κ \kappa κ times the standard deviation of that gain.
The second condition of Eq. (24) limits truncation where the core carries almost no information about the true phase. All scores then scatter about zero, and their maximum over the K K K sizes can exceed the score of the whole ensemble by more than ζ \zeta ζ by chance. In the limiting case of a core with zero coherence, every term y k ( n , r ) y_k(n,r) y k ( n , r ) has zero expectation and different terms are uncorrelated, so that the standard deviation of a k a_k a k is estimated by
σ k = ∣ I ∣ − 1 ( ∑ ( n , r ) ∈ I y k ( n , r ) 2 ) 1 / 2 , (25) \sigma_k = \left|\mathcal{I}\right|^{-1}\left(\sum_{(n,r)\in\mathcal{I}} y_k(n,r)^2\right)^{1/2}, \tag*{(25)} σ k = ∣ I ∣ − 1 ( n , r ) ∈ I ∑ y k ( n , r ) 2 1/2 , (25)
and a max a_{\max} a m a x must exceed λ σ k max \lambda\sigma_{k_{\max}} λ σ k m a x , with the noise multiplier λ = 3 \lambda= 3 λ = 3 . With ANS, the gate on the PCI (below) skips the truncation step at the majority of such centers, so that the choice of λ \lambda λ changes the error little (Table S20 of the supplement).
Table S20. Variants of the truncation step and of FaSHPS, each with the PCI-scaled window. Each entry is the phase linking error over all centers of the evaluation scenes relative to that of the same method with ANS and the selected setting, as a ratio of the weighted geometric means on the same centers. The rows give the noise multiplier λ = 0 \lambda= 0 λ = 0 , 1, 2, 4 and 5 with the selected margin factor κ \kappa κ , the members ordered by their SPC with the center instead of the ellipse order, the ensemble grown in this SPC order up to the candidate size of largest Fisher information ∣ E ∣ Q ‾ / ( 1 − Q ‾ ) |\mathcal{E}|\,\overline{Q}/(1-\overline{Q}) ∣ E ∣ Q / ( 1 − Q ) , and the variant of FaSHPS of § 3.2 that also accepts candidates outside the guard set whose SPC with the center exceeds the 99th percentile of S p q S_{pq} S pq in clutter with zero coherence. A dash marks a variant that was not run.
The rule (Eq. (24) ) has no stated significance level, since the terms of the score are not independent where the core is coherent and the maximum is taken over the sizes.
Gate on the PCI. The gate is a one-sided test of η p \eta_p η p against η 0 \eta_0 η 0 (§ 2.2.3 ). In homogeneous clutter with zero coherence it passes 1 % of the pixels, by the definition of η 0 \eta_0 η 0 . At a center below η 0 \eta_0 η 0 the PCI shows no coherence, and a truncation would respond mostly to noise. In ANS, the truncation step is therefore applied only at centers with η p ≥ η 0 \eta_p \ge\eta_0 η p ≥ η 0 . At the other centers the ensemble of the SHP selection in B p \mathcal{B}_p B p is returned, since B p ′ = B p \mathcal{B}'_p = \mathcal{B}_p B p ′ = B p there. In ANS the SHP selection and the truncation step are otherwise unchanged.
Relation to existing adaptive neighborhoods Three properties distinguish the truncation step from the adaptive neighborhoods and the nonlocal estimators cited in § 2 . First, the reference is the center. The adaptive multilook of Pepe et al. [15 ] , whose agreement measure is of the same kind as Eq. (23) (§ 2.1.2 ), judges every pixel against the multilook of a fixed window, which no longer describes the center where the phase varies within the window. NL-SAR (Deledalle et al. (2015)) selects among nonlocal estimates by their equivalent number of looks, so that phase differences enter only through the similarity of the patches. The truncation step compares the multilooks of nested subsets with the phases of the core around the center and thus responds directly to a bias of the multilooked phases relative to the center. Second, the decision pools all acquisitions and requires neither a phase model nor a separate noise estimate, whereas the ICI rule (Katkovnik et al. (2002)) with local polynomial fits (Katkovnik et al. [8 ] ) compares estimates of one interferogram at a time through confidence intervals. Third, the guard set keeps the pixels whose speckle is correlated with that of the core out of the score, so that shared speckle cannot make a small subset agree with the core. § 14.1 of the supplement applies the local polynomial fit with ICI to every interferogram and compares its error with that of FaSHPS followed by the truncation step at every center and that of FaSHPS+ANS.
Point scatterers
A pixel dominated by a strong, stable scatterer requires a different ensemble, namely the few pixels that carry the response of the same scatterer, which have the highest SPC with the center but need not be the nearest. The methods compared below have no such treatment. Point scatterers are left to the processing of persistent scatterers, and their errors are reported separately (§ 2.3.2 ).
Estimator of the coherence matrix
The coherence matrix of an ensemble E \mathcal{E} E is estimated from amplitude-normalized samples,
Γ ^ = 1 ∣ E ∣ ∑ q ∈ E u q u q H , (26) \widehat{\boldsymbol{\Gamma}} = \frac{1}{\lvert\mathcal{E}\rvert}\sum_{q\in\mathcal{E}} u_q u_q^{H}, \tag*{(26)} Γ = ∣ E ∣ 1 q ∈ E ∑ u q u q H , (26)
whose entries are the complex phase-only coherences, i.e., Eq. (2) without the magnitude. Its entries are the first trigonometric moments of the single-look interferometric phases over the ensemble, so the amplitudes do not enter. Phase linking based on circular statistics uses the same phase-only information (Yao and Balz (2024); Yao et al. (2025)). The phases of Γ ^ \widehat{\boldsymbol{\Gamma}} Γ estimate those of Γ \boldsymbol{\Gamma} Γ , and its magnitudes are reduced by the known, monotone function g g g of Eq. (7) and can be corrected by inverting it. We use Eq. (26) for every method in this paper, because a few bright pixels in a large ensemble dominate the ordinary sample covariance, whereas with Eq. (26) they receive the same weight as any other pixel. In homogeneous Gaussian clutter without texture, however, the ordinary sample covariance is the efficient estimator, and Eq. (26) gives up the amplitude information there. Over the random scenes of the simulations (§ 2.3.1 ), which contain texture and point scatterers, the ordinary sample covariance has a higher error over the distributed centers for every SHP selection with the truncation step or with ANS, most for the full window (Section S3.2 of the supplement).
Computational cost
Per candidate and pass, the amplitude test (Eq. (19) ) requires a few operations, independent of N N N , since the temporal mean amplitudes are computed once per pixel, and the test of CMP a sum over the N − 1 N-1 N − 1 single-reference interferograms. The second pass also needs the reference statistics, i.e., the mean amplitude over the reference set in FaSHPS, and the mean intensity and the multilooked phases over it in CMP. ANS uses the PCI η \eta η , computed once per image pixel (§ 2.2.2 ). FaSHPS and CMP need it for their thresholds in any case (§ 2.2.5 ), and for the full window it is an additional preparation. The truncation step costs considerably more than the PCI and the SHP tests (Table S17 of the supplement). It adds the SPC of the center with the pixels of F p F_p F p within the search window for the ordering (Eq. (20) ), one complex inner product of length N N N each, a sort of the members, R R R running sums of length N N N per member, which give the multilooks of all K K K nested subsets in one pass over the ensemble, and the evaluation of Eq. (23) at K K K sizes, where K K K grows with the logarithm of the ensemble size. The PCI-scaled window reduces the number of candidates at coherent centers, and together with the truncated ensembles it lowers the cost of the coherence matrix itself. § 3.1.4 reports the measured times.
Table S17. Time of the individual steps for one center pixel, or for one image pixel where marked, relative to one sample covariance over the whole search window, for Visp, the slow-deformation test area and the landslide substack (first row, whose absolute time on the test machine, one core of an Intel Xeon E5-2697 v3, was 4.92, 7.35 and 0.79 ms, respectively). The second row, with a quarter of the pixels, shows that the time of the sample covariance is proportional to the ensemble size. FaSHPS preparation is the temporal mean amplitude of each pixel after the removal of outlying amplitudes (Eq. (19) ). All steps are single-threaded C loops compiled with the same settings and checked against the simulation code. The truncation step comprises the multilooked interferograms of the score, its evaluation at the candidate sizes, the jackknife noise and the decision. Median over 150 center pixels per stack, best of three runs each. The NSPC row comes from this separate measurement and differs slightly from the value in Table 4 of the main text.
Simulation study
Design
Questions The simulations ask
how much ANS changes the phase linking error after each SHP selection, and under which local conditions,
what its two parts and the gate contribute,
how much the choice of SHP selection, and the settings of its tests, matter once ANS is applied,
how ANS and the truncation step change the dependence of this error on the size of the search window that the user chooses,
how the results change with fewer acquisitions.
Scenes The benchmark consists of random scenes, each composed of independent components (Table S6 of the supplement). A scene contains one to three clutter classes, each with its own coherence c c c , backscatter, texture and point scatterers. Different kinds of land cover differ in the same properties, but the classes are not modeled on any particular one. The true phase of a scene is the sum of a deformation phase and an independent atmospheric phase screen for each acquisition. The deformation phase can be discontinuous along a fault or at the edge of a region that moves while its surroundings stay still, as a landslide body does. Point scatterers add the phase of their height. The evaluated region contains the centers whose search window lies inside the scene. Figure 4 shows a scene and its components.
Figure 4. One of the random scenes of the TSX-44 configuration and its components, for the interferogram of the last acquisition with the first. (a) Two clutter classes with coherence c = 0.95 c = 0.95 c = 0.95 and c = 0.50 c = 0.50 c = 0.50 (Eq. (27) ). (b) Temporal mean amplitude of the simulated stack, with the point scatterers of both classes. In the coherent class the speckle persists in the temporal mean. (c) Deformation phase, discontinuous along a fault (vertical) and at the edge of a moving region (lower right). (d) Atmospheric phase, unwrapped and on its own scale. (e) True phase, the sum of (c) and (d). Each point scatterer adds the phase of its height above the ground, 0.37 rad per meter of height in this interferogram, i.e., more than one cycle at 20 m. This phase per meter is 3.10 times the standard deviation of the phase per meter over the acquisitions (Table S6 of the supplement). (f) Simulated single-look phase. The dashed frame marks the evaluated region and the circle the search window of one center.
Table S6. Components of the random scenes. Lengths are in window radii except the scatterer layout, which is in pixels. Values given as ranges are drawn uniformly unless stated otherwise.
The deformation phase of the last acquisition relative to the first comes from a single smooth source, which forms a ramp, a bowl, a cone, a trough or a shape in between, depending on its random parameters (Figure 4 c, Table S6 of the supplement). The field is scaled so that its steepest gradient in the evaluated region gives F max F_{\max} F m a x fringes per window diameter (Table 2 gives one fringe per window diameter in displacement and rate), and the deformation grows linearly in time. A window radius is one unit of the normalized distance of Eq. (16) , and a window diameter two. The local fringe density of a center is that of this field at the center, and every scene therefore contains a continuum of local fringe densities up to F max F_{\max} F m a x . In half of the scenes with more than one clutter class, the classes also move separately, so that a class border is a motion border as well. In the other scenes, all classes share the deformation, as a surface with several kinds of cover moves as a whole. The atmospheric phase screens have the power-law spectrum of Kolmogorov turbulence ([4 ] , Eq. (4.7.11)), with the parameters given in Table S6 of the supplement. Each scene has one of three strengths (zero, moderate and strong). The strength is the root mean square (RMS) phase difference across one window diameter in an interferogram of two acquisitions. The moderate strength of 0.3 rad is of the order of what the delay model of Emardson et al. (2003) gives at the window extents. For distances much shorter than 1 km they give an exponent of about 5/6 (their Eq. (33)) with the same coefficient as their model for 10 to 500 km and small height variations (their Eq. (32)), whose value they estimate from distances of 10 km and more. We use this coefficient, with which the two forms agree at 1 km. The strong one of 1 rad is a stress case. The true phase history of a center includes its own atmosphere, since phase linking estimates the total phase.
The coherence of each class follows a one-parameter exponential model (cf. Rocca, 2007, Eq. (3.6)) with c c c between consecutive acquisitions, decaying to c 2 c^2 c 2 with a time constant of ten acquisitions,
∣ γ s t ∣ = c 2 + ( c − c 2 ) e − ( ∣ s − t ∣ − 1 ) / 10 , s ≠ t . (27) \lvert\gamma_{st}\rvert= c^2 + \left(c - c^2\right)e^{-\left(\lvert s-t\rvert-1\right)/10}, \qquad s \ne t. \tag*{(27)} ∣ γ s t ∣ = c 2 + ( c − c 2 ) e − ( ∣ s − t ∣ − 1 ) /10 , s = t . (27)
The images are generated in four steps. Every pixel receives an independent draw from the N N N -dimensional zero-mean complex Gaussian distribution whose coherence matrix is real and given by Eq. (27) with the c c c of its class. The draw is multiplied by the amplitude and texture of the pixel and by its true phase, the complex amplitudes of the point scatterers are added, and every acquisition is filtered with the point spread function of the configuration, a separable raised-cosine spectrum whose bandwidth gives the lag-one speckle correlation of Table 2 . Because the phase is applied before the point spread function, a phase gradient inside a resolution cell reduces the coherence, and the sidelobes of a point scatterer transfer its phase to neighboring pixels.
A small set of elementary scenes isolates single mechanisms: homogeneous clutter at $c = 0.2, 0.5 and 0.95, phase steps of 1, 3 and 6.3 rad at the end of the stack at c = 0.65 c = 0.65 c = 0.65 and 0.95, a phase step of 3 rad combined with an amplitude contrast of 6 dB, and a coherence step between c = 0.65 c = 0.65 c = 0.65 and c = 0.2 c = 0.2 c = 0.2 with equal phase and amplitude. The elementary scenes serve as illustration, are never averaged together with the random scenes, and are reported in full in Table S9 of the supplement.
Table S9. Elementary scenes of Section 4.1, TSX-44, one kind per row: phase linking error relative to the oracle, weighted geometric mean over the centers of the elementary scenes of the row, and their median oracle error. Columns as in Table 3. The lowest value of each row is set in bold.
Configurations and centers Three configurations mirror the real stacks of § 3 in stack size, search window and point spread function (Table 2 ). Each of these configurations, TSX-44, S1-79 and S1-20, has its own random scenes, drawn from the same distribution of components (Table S6 of the supplement) and synthesized on its own pixel grid with its number of acquisitions and point spread function. S1-20 mirrors the landslide of § 3.1 , which is processed as a substack of 20 acquisitions with the settings of this configuration, but the simulated scenes of S1-20 are drawn independently of those of S1-79. In contrast to the real data, no mask is applied in the simulations. In each scene, 40 centers are drawn uniformly over the evaluated region. Ten further centers are drawn from each of four subsets of it if the scene has such pixels, so that a scene has 40 to 80 centers. The subsets are the pixels within 0.25 window radii of a class border, within 0.25 window radii of a discontinuity, on a point scatterer, and in the clutter one to six pixels from a scatterer. Centers that do not lie on a point scatterer are distributed centers, and clutter is defined as in § 2.1.5 . We call the distributed centers of one value of c c c a coherence class. With every center we store its local fringe density, the coherence of its clutter class, its distances to the nearest class border and discontinuity, its texture, whether it lies on a scatterer, and the atmospheric strength of its scene.
Methods Each SHP selection is followed by the estimator (Eq. (26) ) and by the same phase linking step. With γ ^ s t = ∣ γ ^ s t ∣ e i φ ^ s t \hat{\gamma}_{st}=|\hat{\gamma}_{st}|e^{i\hat{\varphi}_{st}} γ ^ s t = ∣ γ ^ s t ∣ e i φ ^ s t the entries of Γ ^ \hat{\mathbf{\Gamma}} Γ ^ , the phase linking estimate θ ^ = ( θ ^ 1 , … , θ ^ N ) \hat{\boldsymbol{\theta}}=(\hat{\theta}_1,\ldots,\hat{\theta}_N) θ ^ = ( θ ^ 1 , … , θ ^ N ) is the maximizer over θ \boldsymbol{\theta} θ of
J ( θ ) = ∑ s , t ∣ γ ^ s t ∣ 2 cos ( φ ^ s t − θ s + θ t ) = ϑ q H ( ∣ Γ ^ ∣ ⊙ Γ ^ ) ϑ q , ϑ n = e i θ n , (28) J(\boldsymbol{\theta})=\sum_{s,t}|\hat{\gamma}_{st}|^2\cos(\hat{\varphi}_{st}-\theta_s+\theta_t)=\boldsymbol{\vartheta}^{qH}\left(|\hat{\mathbf{\Gamma}}|\odot\hat{\mathbf{\Gamma}}\right)\boldsymbol{\vartheta}^{q},\qquad\vartheta_n=e^{i\theta_n}, \tag*{(28)} J ( θ ) = s , t ∑ ∣ γ ^ s t ∣ 2 cos ( φ ^ s t − θ s + θ t ) = ϑ q H ( ∣ Γ ^ ∣ ⊙ Γ ^ ) ϑ q , ϑ n = e i θ n , (28)
where ⊙ \odot ⊙ is the entrywise product. This is the phase triangulation coherence maximization in the form of Even and Schulz (2018, Eq. (15)) with exponent 2, i.e., with the squared coherence magnitude as weight. We maximize J J J by cyclic coordinate ascent and damped Newton iterations from the phases of the leading eigenvector of ∣ Γ ^ ∣ ⊙ Γ ^ |\hat{\mathbf{\Gamma}}|\odot\hat{\mathbf{\Gamma}} ∣ Γ ^ ∣ ⊙ Γ ^ (Section S3.4 of the supplement), with the same solver on the real data.
The main comparison contains the three SHP selections of § 2.2 , the full window, FaSHPS and CMP, each alone and in three variants, with the settings of Table 2 . We write X+S for an SHP selection X in the PCI-scaled window, X+T for X followed by the truncation step at every center, without the gate, and X+ANS for X with ANS. With S, the full window is the whole PCI-scaled window. The supplement reports checks of the design (Section S3.1), the thresholds of the SHP tests (Section S2.2) and further methods and options run under the same protocol (§ 14.1 ).
Errors and reference We compute three errors per center. The matrix error is the RMS, over all acquisition pairs, of the wrapped difference between the phase of the estimated coherence matrix and the true phase difference of the center. The phase linking error is the same for the phase linking estimate and is the main measure. Using all pairs makes it independent of a reference acquisition. The coherence error is the RMS distance between estimated and true complex coherence,
e Γ 2 = 1 M ∑ s < t [ ( ∣ γ ^ s t ∣ − ∣ γ s t ∣ ) 2 ⏟ magnitude + 2 ∣ γ ^ s t ∣ ∣ γ s t ∣ ( 1 − cos ϵ s t ) ⏟ phase, weighted by magnitude ] , (29) e_{\mathrm{\Gamma}}^2=\frac{1}{M}\sum_{s<t}\left[\underbrace{\left(|\hat{\gamma}_{st}|-|\gamma_{st}|\right)^2}_{\text{magnitude}}+\underbrace{2|\hat{\gamma}_{st}||\gamma_{st}|(1-\cos\epsilon_{st})}_{\text{phase, weighted by magnitude}}\right], \tag*{(29)} e Γ 2 = M 1 s < t ∑ magnitude ( ∣ γ ^ s t ∣ − ∣ γ s t ∣ ) 2 + phase, weighted by magnitude 2∣ γ ^ s t ∣∣ γ s t ∣ ( 1 − cos ϵ s t ) , (29)
where ϵ s t \epsilon_{st} ϵ s t is the phase error of pair ( s , t ) (s,t) ( s , t ) , and the true coherence magnitude is the coherence of the class, reduced by the decorrelation that the difference of the local phase gradients of the two acquisitions causes within the resolution cell, and mapped by g g g of Eq. (7) for the estimator (Eq. (26) ). The reduction of the magnitude by Eq. (26) is a property of the estimator and not of the SHP selection, and it is therefore not counted as an error. At a point scatterer no true coherence magnitude is defined, so that the coherence error is given over the distributed centers only.
Errors are compared through a fixed reference, the oracle . For every center it is the nested subset of the whole search window, in the distance order or the ellipse order of the truncation step (§ 2.2.6 ), at sizes from 25 pixels upward in steps of a factor 1.5 \sqrt{1.5} 1.5 , with the lowest phase linking error. Its size is the oracle size and its error the oracle error
e p ⋆ = min o ∈ { d , e } min n ∈ K p e P L ( P p o ( n ) ) , K p = { round ( 25 ⋅ 1.5 j / 2 ) < n p : j = 0 , 1 , … } ∪ { n p } , (30) e_p^\star=\min_{o\in\{d,e\}}\min_{n\in\mathcal{K}_p}e_{\mathrm{PL}}\left(\mathcal{P}_p^o(n)\right),\qquad\mathcal{K}_p=\left\{\operatorname{round}\left(25\cdot1.5^{j/2}\right)<n_p:j=0,1,\ldots\right\}\cup\{n_p\}, \tag*{(30)} e p ⋆ = o ∈ { d , e } min n ∈ K p min e PL ( P p o ( n ) ) , K p = { round ( 25 ⋅ 1. 5 j /2 ) < n p : j = 0 , 1 , … } ∪ { n p } , (30)
where P p o ( n ) \mathcal{P}_p^o(n) P p o ( n ) is the set of the first n n n pixels of the search window of p p p in the distance order (o = d o=d o = d ) or the ellipse order (o = e o=e o = e ), n p n_p n p the number of pixels of the window and e P L e_{\mathrm{PL}} e PL the phase linking error of an ensemble. The oracle requires the true phase and is not a method. Every error is divided by the oracle error at the same center, and we call this ratio the error relative to the oracle. Since the reference does not depend on the methods, adding, removing or retuning one method changes no other method’s ratios. A ratio below one means that an SHP selection found a better ensemble than the best nested subset that the oracle tries. At centers with zero coherence the oracle picks the nested subset whose noise happens to lie closest to the true phase, so that a ratio above one there does not indicate room for improvement.
Summaries are weighted geometric means of this ratio over a stated set of centers. The weight of a center is the uniform density over the evaluated region divided by the density with which the draws of its scene sample its position, normalized to a sum of one over the centers of the scene. The weighted centers of a scene thus represent its area, and every scene counts equally. Centers on point scatterers, for example, make up 12 to 13 % of the centers but 0.4 % of the weight in each configuration. A comparison of two methods is given as the ratio of their means. This ratio equals the weighted geometric mean of the ratio of their errors and therefore does not depend on the reference. Section S3.6 of the supplement gives confidence intervals of these ratios over all centers from a bootstrap resampling of the scenes, and the main results with the arithmetic mean and the root mean square of the error in place of the geometric mean. An average over all centers depends on the distribution of conditions in the scenes, and in particular on the coherence classes, which are equally frequent here. We therefore give the results against local fringe density and coherence, and we do not rank methods on an overall average alone.
Protocol Three disjoint sets of scenes are used, each generated in batches from its own random seeds, from which every random element of a scene and of its centers derives. The calibration scenes, homogeneous clutter at eight coherences from 0 to 0.95 in each configuration, give the second-pass threshold of FaSHPS and the thresholds of CMP over the PCI (Section S2.2 of the supplement) and, at zero coherence, the null quantile η 0 \eta_{0} η 0 (Section 3.3). On the tuning scenes (three batches, 600 random scenes and 39,950 to 40,520 centers per configuration), ρ min \rho_{\min} ρ m i n , κ \kappa κ and λ \lambda λ are tuned in turn by a criterion fixed before the tuning runs. That criterion is the weighted geometric mean of the phase linking error relative to the oracle over all centers, averaged over the SHP selections with ANS and over the configurations. Section S3.1 of the supplement gives the rules of this tuning and its results. The evaluation scenes (five batches, 2000 random scenes and 134,480 to 135,460 centers per configuration, and 55 elementary scenes per configuration) give the results of § 2.3.2 . No setting is tuned on them. The windows of lowest error in Table S8 of the supplement are identified in hindsight, and the variants and the fixed constants of Section S5 of the supplement are compared on them.
Reported values In the results of the simulations, a value or a range refers to all centers, the three SHP selections and the configurations TSX-44 and S1-79, unless the sentence or the parentheses after it name other centers, SHP selections or configurations. A range gives the smallest and the largest value over them. S1-20 is reported in the tables and in the paragraph on the smaller stack (§ 2.3.2 ). Values in the text are computed from the unrounded summaries and can therefore differ from values computed from the rounded entries of the tables.
Results
Where ANS lowers the error ANS lowers the error after every SHP selection, and most of the reduction comes from the truncation step, which acts most strongly in coherent fringes. Over all centers it lowers the error by 21 to 31 % relative to the SHP selection alone (Table 3 ). Without the truncation step the error grows steeply with the fringe density in coherent clutter (Figure 5 ). At two or more fringes per window diameter and c ≥ 0.5 c \ge0.5 c ≥ 0.5 , the truncation step lowers the error of the full window 2.57 to 8.46 times, and the weighted median of its error from at least 42 ∘ 42^\circ 4 2 ∘ to at most 10 ∘ 10^\circ 1 0 ∘ . In zero-coherence clutter it changes the error little, as intended, since there the fringes carry no information. Below 0.25 fringes per window diameter its reduction of the error of the full window grows with the coherence, because atmospheric screens, class borders and discontinuities make a smaller ensemble preferable even without deformation fringes.
Figure 5. The truncation step removes most of the growth of the phase linking error of FaSHPS with the fringe density in coherent clutter. Phase linking error of FaSHPS relative to the oracle (Eq. (30) ) against the local fringe density of the deformation between the first and the last acquisition, for four coherence classes in the configurations TSX-44 and S1-79 (44 TerraSAR-X and 79 Sentinel-1 acquisitions), weighted geometric mean over the distributed centers of the evaluation scenes in each fringe density bin with at least 30 centers. Colors show the variant, FaSHPS alone (blue), with the PCI-scaled window (+S, orange), with the truncation step (+T, cyan) and with both (+ANS, red). Without the truncation step the error grows steeply with the fringe density in coherent clutter, also within the PCI-scaled window. The full window and CMP, which are not drawn, lie within 15 and 8.1 % of FaSHPS, respectively, in every bin and variant.
Table 3. Phase linking error relative to the oracle (weighted geometric mean, Eq. (30) ) and weighted median oracle error, for the distributed centers of each coherence class, the distributed centers with at least one fringe per window diameter, the centers on point scatterers and all centers (evaluation scenes, amplitude-normalized estimator). The three blocks give the configurations TSX-44, S1-79 and S1-20 (44 TerraSAR-X, 79 and 20 Sentinel-1 acquisitions). Full stands for the full window, S marks the PCI-scaled window and T the truncation step at every center. ANS combines both and applies the truncation step only where η p ≥ η 0 \eta_p \ge\eta_0 η p ≥ η 0 (§ 2.2.3 and § 2.2.6 ). The lowest value of each row is set in bold, except in the rows over all centers.
Where nothing calls for truncation, ANS can raise the error. Per center it raises the error by more than 5 % relative to the SHP selection alone at 13 to 17 % of the area, mostly in coherent clutter with few fringes, where the error is a few degrees. In the homogeneous elementary scenes of TSX-44, where ensembles are truncated only by chance, the truncation step alone raises the 90th percentile of the phase linking error of the full window by less than 4 ∘ 4^\circ 4 ∘ . ANS raises the error more than the truncation step alone in most of these scenes (Section S3.9 of the supplement). The coherence error and the matrix error follow the phase linking error (Section S3.1.1 and Table S10 of the supplement).
Table S10. Matrix error relative to the matrix error at the oracle size, weighted geometric mean over the distributed centers of three coherence classes and over all centers (evaluation scenes, amplitude-normalized estimator), and the weighted median matrix error at the oracle size. The oracle is chosen by the phase linking error. Columns as in Table 3 , without the PCI-scaled window alone (+S).
Contributions of the two parts The truncation step does most of the work of ANS. Alone, it lowers the error by 22 to 30 %, and the PCI-scaled window by 14 to 20 % (Table 3 ). Together they lower it about as much as the truncation step alone over all centers, and by a further 2.2 to 4.8 % at one fringe or more per window diameter. The PCI-scaled window thus helps where the window contains coherent fringes or an atmospheric screen (§ 13.3.10 of the supplement). Where the clutter is uniform over regions much larger than the search window, as in most elementary scenes, the PCI-scaled window raises the error, since it removes looks that need not be excluded, although the errors there are small in degrees. In the random scenes the truncation step alone is marginally better than ANS in weakly coherent clutter (Table 3 ). On the real stacks the processing with ANS takes 14 to 32 % less time than with the truncation step alone (§ 3.1.4 ). The gate keeps the truncation step from acting on noise. In the coherence class c = 0 c = 0 c = 0 of the random scenes it skips the truncation step at 74 to 90 % of the distributed centers. It rarely passes those far from point scatterers and class borders, and often those near them (Section S3.7 of the supplement), and ANS leaves the error of the SHP selection nearly unchanged.
SHP selections and their tests ANS improves every SHP selection, and with ANS the three end within a few percent of each other (Table 3 ). FaSHPS and CMP keep an advantage of 0.7 to 3.9 % over the full window with ANS over all centers, more than the sampling noise of the scenes except for CMP in TSX-44 (Section S3.6 of the supplement). Without either step they accept most of the search window, FaSHPS about four fifths and CMP about three quarters of it, at the backscatter contrasts of the scenes, whose classes differ by at most 6 dB. In zero-coherence clutter the full window has a lower error than FaSHPS and CMP, because the SHP tests exclude coherent neighbors of other classes, which carry information on the true phase where the classes move together. On point scatterers the test of CMP brings the largest reduction, with 0.40 to 0.60 times the error of FaSHPS, alone and with ANS (all three configurations).
A different significance level of the SHP tests or their original thresholds in place of those that depend on the PCI change the error by at most 6.7% (Table S4 of the supplement), much less than adding ANS does. A level of 0.1% gives a slightly lower error than 1% here, because the scenes contain no narrow structures, such as roads a few pixels wide, at which the additional candidates of a lower level would raise the error. We therefore keep 1%.
Table S4. Checks of the design: phase linking error over all centers of the evaluation scenes relative to that of the same method with ANS, ratio of the weighted geometric means on the same centers. Each row changes one element, namely the gated truncation step on the search window without the PCI-scaled window, the distance order in place of the ellipse order, a fixed margin ζ = 1.5 / N \zeta=1.5/N ζ = 1.5/ N in place of κ σ Δ a \kappa\sigma_{\Delta a} κ σ Δ a , the null quantile η 0 \eta_0 η 0 at other percentiles, a window that shrinks to 25 pixels already at the PCI of c = 0.95 c=0.95 c = 0.95 without the truncation step, other significance levels of the SHP tests, and their original thresholds (Section S3.1). The original threshold with the SHP selection alone is compared with the same SHP selection alone. A dash marks a variant that was not run or that does not apply to the full window, which has no SHP test.
Size of the search window Without either step, every SHP selection has its lowest error at a middle window and 13 to 31% more at the largest (Figure 6 , Table S8 of the supplement). With ANS the largest window has the lowest error, 9.6 to 11% below that of the same SHP selection without either step at its best window, so that a window large enough for the decorrelated parts of a scene adds little error at the others. The search windows of Table 2 are those used on the real stacks, where decorrelated areas require many looks. With the smaller windows the truncation step alone has a lower error than ANS. ANS catches up with it at the two largest windows in TSX-44 and at the largest in S1-79. At the windows of Table 2 the processing with ANS takes less time than with the truncation step alone on the real stacks (§ 3.1.4 ). Over all centers the size of the window changes the error more than the choice of SHP selection. The error differs by at least 13% between the best and the worst tested window and by at most 4.6% between the SHP selections at one window (without either step, with the truncation step alone and with ANS). On point scatterers it is the other way round. There, without either step, the choice of SHP selection changes the error by at least 129% and the window by at most 27%.
Figure 6. With ANS the largest window, which decorrelated areas need, has the lowest error of the tested windows for every SHP selection. Phase linking error of each window relative to the oracle in the largest window, weighted geometric mean over all centers of the evaluation scenes, against the search window (azimuth × \times × range, ellipse inscribed in the stated rectangle), in the configurations TSX-44 and S1-79 (44 TerraSAR-X and 79 Sentinel-1 acquisitions). Colors show the variant, the SHP selection alone (blue), with the truncation step (+T, cyan) and with ANS (+ANS, red), and line styles the SHP selection, the full window (dotted), FaSHPS (solid) and CMP (dashed). The values are in Table S8 of the supplement. Each SHP selection is drawn separately. Without either step the error is lowest at a middle window, and with ANS at the largest. At the smaller windows the truncation step alone has a lower error than ANS, which catches up with it at the largest.
Table S8. Phase linking error relative to the oracle in the largest window, weighted geometric mean over all centers of the evaluation scenes, for four search windows (azimuth × \times × range, ellipse inscribed in the stated rectangle). The largest is the search window of Table 2 . With ANS the PCI-scaled window shrinks from the window of each row. Settings and thresholds are those of the main comparison. The columns are those of Table 3 without the PCI-scaled window alone (+S).
Smaller stack The smaller stack, S1-20, differs from the larger one, S1-79, only in the number of acquisitions. With 20 acquisitions the PCI is noisier, and its null quantile is η 0 = 0.100 \eta_0 = 0.100 η 0 = 0.100 against 0.050 with 79 (Table 2 ). The SHP selections then differ slightly more than in S1-79, and the advantage of the largest window with ANS persists (Table 3 and Table S8 of the supplement). At high coherence, the original thresholds of FaSHPS reject fewer homogeneous candidates in excess with fewer acquisitions. In homogeneous clutter at c = 0.95 c = 0.95 c = 0.95 they reject 10% of the candidates in the smaller stack and 27% in the larger one, whereas the thresholds that depend on the PCI reject at most 1.8% at every simulated coherence (Figure 3 ).
Experiments on TerraSAR-X and Sentinel-1 data
The real stacks show what the mask, the PCI and ANS do where no true phase is known. § 3.1 describes the datasets. We then check the speckle correlation that sets the correlated-noise floor of NSPC (§ 3.1.1 ), compare the mask with layover and shadow maps (§ 3.1.2 ) and show how ANS changes the phase-linked maps (§ 3.1.3 ). § 3.1.4 reports the runtime on these stacks, and § 3.1.5 explains why we report no quantitative comparison of the SHP selections on real data.
Datasets
The TerraSAR-X stack consists of 44 coregistered acquisitions over Visp, Switzerland, in StripMap mode. The acquisitions were taken in the summer months between 30 June 2017 and 26 September 2022. We use a scene of 1000 × \times × 1000 pixels. The topographic phase was simulated from the terrain model swissALTI3D at 0.5 m spacing (Federal Office of Topography swisstopo, 2019) and removed. A layover and shadow map in radar geometry, derived from the same model with the geocoding of the GAMMA software (Wegmüller, 1999; Frey et al., 2013; Werner et al., 2000}), is used for comparison.
The Sentinel-1 site is a section of the Hunza valley in the Karakoram (Gilgit-Baltistan, Pakistan), about 55 km long and 15 km wide, along the Hunza river and the Karakoram Highway from Rahimabad to Shishkat. Its arid slopes of bare rock and debris remain coherent over years, its extreme relief causes extended layover and shadow, and landslide inventories along the highway, compiled or updated with InSAR, list hundreds of landslides (Hussain et al., [5 ] ; Su et al., [18 ] ) (Section S4.4 of the supplement). The stack consists of 79 Sentinel-1A acquisitions in interferometric wide-swath mode between 5 January 2023 and 20 December 2025, coregistered and deramped with the GAMMA software (Wegmüller et al., [22 ] ) (Section S4.4 of the supplement). No spectral filtering was applied. The topographic phase was simulated from the Copernicus DEM at one arc-second posting (European Space Agency and Airbus, 2022), a surface model derived from TanDEM-X data, and removed. The geometric layover and shadow map used below is derived from the same model. The coregistered SLC data is oversampled by a factor of two in range. We use five test areas of 300 × 2700 pixels (Table S14 of the supplement). They cover the area with the strongest decorrelation of the five, a slope with slow deformation, the slow-deformation test area, a landslide with fast deformation, the fast-deformation test area, an area with extended layover and shadow, the layover and shadow test area, and an area that remains coherent over the whole period, the long-term coherent test area. The test area with the strongest decorrelation is centered on Maiun, and we refer to it as the Maiun test area. It decorrelates over long intervals and keeps coherence over short ones. The landslide of the fast-deformation test area is coherent only over short intervals, so that we additionally use a substack of the 20 acquisitions between 25 November 2023 and 8 September 2024, with the mask computed from the 79 acquisitions of the test area.
Table S14. Approximate center coordinates of the Sentinel-1 test areas and their use in this paper.
Every SHP selection and the truncation step use the settings of Table 2 , with the thresholds of FaSHPS and CMP and the null quantile η 0 \eta_0 η 0 of the matching simulated configuration, TSX-44 for Visp, S1-79 for the Sentinel-1 test areas and S1-20 for the landslide substack (§ 2.3.1 ). For CMP, the decay constant of the coherence model is fitted to each stack with the estimator used for the simulated scenes and the temporal baselines of its acquisitions, and its thresholds are tabulated for this constant in simulated clutter of the configuration. As in the simulations, point scatterers receive no separate treatment.
Speckle correlation and correlated-noise floor
On TerraSAR-X the speckle correlation estimated from the stack agrees with the prediction from the processing parameters. The processing uses this prediction (§ 2.1.5 ). Only pixels at which the data shows no detectable coherence enter the estimate, i.e., the pixels whose PCI (§ 2.2.2 ) lies below its null quantile (§ 2.2.3 ), so that their pairs at lag one along each image axis show no signal beyond the speckle. The mean Q ˉ 1 \bar{Q}_1 Q ˉ 1 of Q ρ q Q_{\rho q} Q ρq over these pairs estimates g ( ∣ ρ ∣ ) 2 g(|\rho|)^2 g ( ∣ ρ ∣ ) 2 by Eq. (10) . We solve
g ( ∣ ρ ∣ ) 2 = Q ˉ 1 (31) g(|\rho|)^2 = \bar{Q}_1 \tag*{(31)} g ( ∣ ρ ∣ ) 2 = Q ˉ 1 (31)
for ∣ ρ ∣ |\rho| ∣ ρ ∣ . § 13.1.3 of the supplement describes the extension of the estimate to the other lags and the further selection of the pairs, which avoids a bias of Q ˉ 1 \bar{Q}_1 Q ˉ 1 by a subtracted model phase and by the brightness of the surface. On the Visp stack the estimate is 0.68 in azimuth and 0.62 in range, compared with the predicted 0.695 and 0.606, respectively.
Adjacent pixels share speckle, so that their SPC lies far above that of independent pixels and falls quickly with the lag (Figure 7 a). The NSPC histogram of the Visp scene peaks at 0.43, where simulated clutter with zero coherence and the speckle correlation of this stack has its median, whereas clutter of independent pixels would have its median at 0.05 (Figure 7 c). Most of the scene thus agrees with its neighbors only to the extent that shared speckle produces.
Figure 7. Adjacent pixels share speckle, so that their SPC lies far above that of independent pixels, and the NSPC histogram has its main mode at this correlated-noise floor. Spatial correlation in the Visp stack (1000 × 1000 pixels, N = 44 N = 44 N = 44 ). (a) Mean corrected SPC versus lag in azimuth and range. The dotted line marks zero, the expected value for independent pixels with zero coherence. (b) NSPC, the mean of the square root of the clipped corrected SPC over the eight direct neighbors (Eq. (14) ), after topographic phase removal. The mean amplitude of the same scene is shown in Figure 8 a. (c) NSPC histogram compared with simulated zero-coherence clutter with the measured speckle correlation and with independent pixels.
On Sentinel-1, the estimated correlation in azimuth lies below the prediction. Equation (11) with the Sentinel-1 processing parameters (Table S11 of the supplement) predicts 0.82 at range lag one and 0.66 in azimuth. The estimator gives 0.50 to 0.61 in azimuth in the Maiun, fast-deformation and slow-deformation test areas and in the layover and shadow test area. The correlation of the intensities of neighboring pixels over time, in contrast, measures ∣ ρ ∣ 2 |\rho|^2 ∣ ρ ∣ 2 , and its square root is close to the prediction over bright surfaces and lower in azimuth over dark surfaces, as expected for thermal noise, which is less correlated in azimuth than speckle (§ 13.1.3 of the supplement). As on TerraSAR-X, the processing uses the prediction.
Mask versus layover and shadow
On both sensors the mask covers a large part of the layover, and on Sentinel-1 also of the shadow. On Visp it also flags pixels in the shadow of buildings and dark pixels near bright scatterers, which the terrain model map does not contain. We set h h h by inspection of the NSPC histogram of each stack, at h = 0.25 h = 0.25 h = 0.25 on TerraSAR-X and h = 0.30 h = 0.30 h = 0.30 on Sentinel-1. Simulated clutter with zero coherence falls below h = 0.30 h = 0.30 h = 0.30 at none of its 230,010 pixels with 79 acquisitions. On Visp, 3.7 % of the valid pixels lie below h h h . They form the layover band on the valley slope and compact patches in the town section of the scene (Figure 7 b and Figure 8 ).
Figure 8. NSPC mask (mask threshold h = 0.25 h = 0.25 h = 0.25 , after topographic phase removal) compared with layover and shadow maps. Over the whole scene almost half of the masked pixels lie in layover of the terrain model. In the town section most of them lie in layover or shadow of the surface model, which contains the buildings. Top: whole scene (1000 × 1000 1000 \times1000 1000 × 1000 pixels). (a) Mean amplitude. (b) Mask by class of the terrain model swissALTI3D. Bottom: town section (dashed box in a and b). (c) Mean amplitude. (d) Mask by class of the surface model swissSURFACE3D Raster. Table S12 of the supplement gives the shares of the masked pixels in each class and of each class that is masked. Source of the elevation models: Federal Office of Topography swisstopo.
In layover, the low NSPC arises from the removal of the topographic phase. The mask covers 56 % of the layover of the terrain model on Visp and 31 to 63 % of the layover in the Sentinel-1 test areas, against 6.3 to 13.8 % of the pixels in neither class there (Figure 9 , Table S15 of the supplement). Before the removal it covers less than 5 % of the layover on both sensors (§ 14.0.2 and S4.6 of the supplement). A layover pixel collects echoes from several heights, and the simulated topographic phase, computed for one height per pixel, changes strongly between adjacent pixels where the phase of the data, dominated by shared speckle and mixed returns, does not change with it. The mask thus flags pixels where the topographic correction applied to the stack is inconsistent with the local phase structure of the data.
Figure 9. On Sentinel-1 the NSPC mask covers a large part of the layover and of the shadow. Test area with layover and shadow (N = 79 N = 79 N = 79 ). (a) Mean amplitude. (b) NSPC. (c) NSPC mask at the mask threshold h = 0.30 h = 0.30 h = 0.30 compared with the geometric layover and shadow map, in the colors of Figure 8 . Each pixel is colored by its class in the map, layover in blue and shadow in orange, dark where the NSPC mask flags it and light where it does not. Masked pixels outside both classes are green, the other pixels light gray. For display, (a) and (b) are reduced to a seventh of their width with a Lanczos filter on the color image, and (c) shows the class with the largest Lanczos weight. The range axis is drawn compressed by a factor of seven, which approximates the aspect ratio of the test areas on the ground.
Table S15. NSPC mask of the Sentinel-1 test areas (after topographic phase removal, on the oversampled grid) against the layover and shadow map, by class of the map: share of the pixels of the test area in the class (Share), median NSPC of the pixels of the class (NSPC), and share of the pixels of the class whose NSPC lies below h = 0.30 h = 0.30 h = 0.30 , i.e., that the mask removes (Masked). The long-term coherent test area is not part of this comparison (Table S14 ).
A large part of the shadow is masked on Sentinel-1, and hardly any of the shadow of the terrain model on Visp. There the mask covers 3.9 % of this shadow, where NSPC stays near the correlated-noise floor, as expected for clutter with zero coherence. In the Sentinel-1 test areas it covers 33 to 61 % of the shadow. Thermal noise, which dominates there, is less correlated in azimuth than speckle (§ 3.1.1 ), so that its simulated NSPC has a lower median than that of zero-coherence clutter, closer to the threshold, though still above it at most pixels. The NSPC of shadow lies lower still (Table S15 of the supplement) and goes with its low intensity (Section S4.6 of the supplement). The masked share of shadow is nearly the same
before topographic phase removal, so that the low agreement in shadow does not arise from the correction.
On Sentinel-1 the mask relies on the range oversampling. On the native range grid, i.e., with every second range sample, simulated clutter with zero coherence has a median NSPC of 0.32 instead of 0.54 (S1-79). A threshold at the 0.1 % quantile of NSPC in this clutter then masks about twice as many of the pixels in neither class as h = 0.30 h = 0.30 h = 0.30 on the oversampled grid, in each of the four test areas used for the mask, all but the long-term coherent test area (Section S4.6 and Table S14 of the supplement).
The mask also flags pixels outside the layover and shadow of the terrain model. In the town section of Visp, a layover and shadow map derived in the same way from the surface model swissSURFACE3D Raster ((Federal Office of Topography swisstopo, 2022)), which includes buildings, trees and other objects above the ground, places 63 % of the masked pixels in shadow, which covers 31 % of the area there (Figure 8 d). Over the whole scene, 19 % of the masked pixels lie outside both classes of the surface model map. They mark disagreement with the neighbors that neither map describes, in particular near bright scatterers, since dark pixels are masked more often the closer they lie to a bright pixel (§ 14.0.2 of the supplement). The sidelobes of a bright scatterer alternate in sign from one range sample to the next beyond the main lobe, so that where a sidelobe is comparable to the clutter, two adjacent pixels receive the scatterer with opposite signs and the clutter speckle with the same sign, and the two contributions to their relative phasor cancel. A simulation of point scatterers in clutter with zero coherence, with the generator and the point spread function of TSX-44, supports this role of the sidelobes (§ 14.0.3 of the supplement). On Sentinel-1 the region of low NSPC is wider than the geometric shadow, and thin lines of low NSPC appear that the geometric map does not contain (Figure 9 b and Figure 9 c). The mask therefore complements geometric layover and shadow maps, also those derived from a surface model.
Phase-linked maps
ANS changes the phase-linked maps mainly where the PCI is high. It preserves the fringes in the slow-deformation test area and in part on the landslide, and keeps large ensembles where the coherence is low.
The maps are computed with the NSPC mask (§ 2.2.1 ), and masked pixels are shown in gray. On Sentinel-1 the search window is the ellipse inscribed in 25 × 225 pixels (azimuth × range, 4437 pixels), of comparable extent in both ground directions. FaSHPS and CMP accept fewer candidates on the real stacks than in the simulations, on Visp 42 % and 19 %, respectively, of the unmasked pixels of the 101 × 101 search window on average. For FaSHPS the difference arises mainly at centers with a low PCI, where the mean amplitudes spread over the search window about as much as in the simulated clutter with strong texture or more. On Visp CMP accepts fewer candidates than in the simulations at every PCI, since the mean intensity of the candidates differs from that of its reference set more than in the simulated scenes (Section S4.8 of the supplement).
Visp has no dense fringes, and the decorrelation over its forested and vegetated parts is strong. Figure 10 shows the interferogram between the first and the last acquisition (single look in Figure 10 a). With a search window of 31 × 31 pixels, the full window yields a patchwork of phase values over vegetation (Figure 10 b). With 101 × 101 pixels the SHP selections give a continuous field (Figure 10 e to Figure 10 g). A continuous field alone does not show that its phase is correct, since a large ensemble also smooths noise. We use the larger window because in the simulations the largest tested window has the lowest error at centers with zero coherence, for the SHP selections alone, with the truncation step and with ANS (Section S3.8 of the supplement), and with ANS also over all simulated centers (Figure 6 ). The advantage at zero coherence carries over to real data only where the decorrelated pixels move with the coherent pixels within the window.
Figure 10. On Visp, which has no dense fringes, ANS changes the phase-linked maps mainly in the town section and along the band of high η \eta η in the north, where it shrinks the search window and truncates the ensembles. Phase-linked differential interferogram of the Visp scene between the first and the last acquisition (30 June 2017 to 26 September 2022). With the small search window the full window gives a patchwork over vegetation, and with the large one the three SHP selections give continuous fields, the smoothest with the full window. The estimate is computed for every unmasked pixel of the area shown, and pixels masked by NSPC (h = 0.25 h = 0.25 h = 0.25 ) are shown in gray. The search windows are the ellipses inscribed in the stated rectangles. (a) Single-look interferogram. (b) Full window with a search window of 31 × 31 31 \times31 31 × 31 pixels. (c) PCI η \eta η , drawn on [ 0 , 0.6 ] [0, 0.6] [ 0 , 0.6 ] . (d) Ambiguity coefficient of FaSHPS+ANS (Heimpel et al., 2026). (e) to (g) Full window, FaSHPS and CMP with a search window of 101 × 101 101 \times101 101 × 101 pixels. (h) Ensemble size of FaSHPS. (i) to (l) The same with ANS, (h) and (l) on a common scale. (m) to (p) Close-ups of the dashed box, which contains roads crossing the forest: mean amplitude and the three SHP selections with ANS. FaSHPS+ANS keeps the roads as lines of distinct phase and CMP+ANS keeps parts of them, while the full window blends them with the surrounding forest. The labels give the median ensemble size.
On Visp ANS changes little except in the town and along a band of high η \eta η in the north. At the 61 % of the centers with η < η 0 \eta< \eta_0 η < η 0 , it changes neither the search window nor the ensemble. About half of the truncations lie in the town section, which holds a fifth of the centers and where η \eta η is higher than elsewhere (Figure 10 c). There the median ensemble size of FaSHPS falls to about a fifth with ANS, against about two thirds over the whole scene (Figure 10 h and Figure 10 l). In the town and along the band in the north, the phase linking estimate of the SHP selections with ANS varies more from pixel to pixel than without it (Figure 10 e to Figure 10 g and Figure 10 i to Figure 10 k). Whether this variation is the motion of individual structures or the noise of the smaller ensembles cannot be decided from this data. Elsewhere the maps with and without ANS are nearly the same.
The close-ups (Figure 10 m to Figure 10 p) show roads crossing the forest. With ANS, FaSHPS keeps a road as a line of distinct phase, CMP keeps parts of it, and the full window blends it with the surrounding forest. Figure 10 d shows the ambiguity coefficient of Heimpel et al. (2026) for FaSHPS+ANS. It compares the maximum of Eq. (28) with the largest value that a solution orthogonal to the maximizer attains, relative to the value expected for noise (their Eq. (3.22)). We use their dampening constant of 1/3 and fit the noise value as in their Eq. (3.13) for the weights of Eq. (28) . It is close to one where the best orthogonal solution attains no more than the noise value and the maximum lies clearly above that value, as on the roads, in the town and along the railway, and low over most of the forest, consistent with a road being a target of its own. The coefficient describes how distinct the maximum is for a given ensemble. It does not show whether the ensemble is homogeneous.
The slow-deformation test area represents the case for which ANS was designed. The slope remains coherent over the three years and fringes accumulate on it with time, so that the phase is flat within the window for short intervals and varies within it for long ones. In the interferogram of Figure 11 , the fringe in the lower part of the slope can be recognized in the single look (Figure 11 a). In the upper part of the slope the full window, FaSHPS and CMP yield broad patches of nearly constant phase, separated by abrupt phase jumps (Figure 11 d to Figure 11 f). With ANS the phase there forms narrow bands along the slope (Figure 11 g to Figure 11 i). Where the estimate of FaSHPS+ANS forms these bands, the single look of the same interval, averaged over 3 × 27 pixels, has a median of 1.6 fringes per window diameter, so that the fringes there are dense relative to the search window (§ 14.0.7 of the supplement gives further values). The single-look interferograms of the two halves of the interval, averaged over 3 × 27 pixels, show bands of the same orientation (Figure 11 j and Figure 11 k), and, less clearly, so does that of the whole interval with the same average (Figure 11 b). The gate passes almost every center of this test area (Figure 11 c). Where FaSHPS+ANS forms the bands, the truncation step cuts most ensembles, so that they are much smaller there than elsewhere
!Twelve-panel phase, PCI, and ensemble-size maps with panels (a) Single look, (b) 3 × 27 3 \times27 3 × 27 average, (c) PCI η \eta η , (d) Full window, (e) FaSHPS, (f) CMP, (g) Full window+ANS, (h) FaSHPS+ANS, (i) CMP+ANS, (j) First half, 3 × 27 3 \times27 3 × 27 average, (k) Second half, 3 × 27 3 \times27 3 × 27 average, and (l) Ensemble size, FaSHPS+ANS; axes labeled range [px] and azimuth [px], with colorbars for phase [rad], PCI η \eta η , and ensemble size [px] bbox=0.19,0.10,0.85,0.71
Figure 11. On the slope of the Sentinel-1 slow-deformation test area, where fringes accumulate over the interval, the SHP selections without ANS give broad patches separated by phase jumps, and with ANS narrow bands along the slope, with smaller ensembles on the coherent slope. Phase-linked differential interferogram between acquisitions 29 and 66 (31 December 2023 to 23 June 2025), N = 79 N = 79 N = 79 . For display, the color image of each map is reduced to a seventh of its width with a Lanczos filter, so that the color is paler where the phases within the filter disagree, and the range axis is drawn compressed by a factor of seven. Pixels masked by NSPC (h = 0.30 h = 0.30 h = 0.30 ) are shown in gray. The search window is the ellipse inscribed in the stated rectangle. (a) Single-look interferogram. (b) The same averaged over 3 × 27 3 \times27 3 × 27 pixels. (c) PCI η \eta η , drawn on [ 0 , 0.6 ] [0, 0.6] [ 0 , 0.6 ] . (d) to (f) Full window, FaSHPS and CMP with a search window of 25 × 225 25 \times225 25 × 225 pixels. (g) to (i) The same with ANS. (j), (k) Single-look interferograms of the first half (acquisitions 29 to 48) and the second half (acquisitions 48 to 66) of the interval, averaged as in (b). (l) Ensemble size of FaSHPS+ANS. The labels give the median ensemble size.
(Figure 11 ). With ANS the median ensemble size of FaSHPS over the test area falls from 3258 to 1218 pixels (Figure 11 e and h).
Figure 12 shows the landslide in the 20-acquisition substack, where the single-look interferograms contain dense fringes (Figure 12 a). Within the lobes, FaSHPS yields smooth bands that are wider than those of the single look, and part of the fringe pattern of the single look is missing (Figure 12 d). With ANS it follows the fringes of the single look in the lower part of the landslide, where the PCI is higher (Figure 12 c). In the upper part, where the PCI is lowest, it recovers fewer of them, and it is noisier at the margins of the lobes (Figure 12 e). Half of the FaSHPS ensembles are truncated, almost a third of these to less than a tenth of the FaSHPS ensemble at the same center, and the median ensemble size falls to less than half (Figure 12 d to f). Between the first and the last acquisition of the substack, 288 days apart, the single look shows no fringes on the landslide (Figure 12 b). Where dense fringes and strong decorrelation over the substack coincide in this way, ANS recovers the fringes only in part, and a longer wavelength would be better suited to this landslide.
Figure 12. Where dense fringes and strong decorrelation over the substack coincide, as on this landslide in the Sentinel-1 fast-deformation test area, ANS recovers the fringes only in part. Phase-linked 12-day interferogram of two consecutive acquisitions in the middle of the 20-acquisition substack. FaSHPS alone smooths the dense fringes into wider bands, and with ANS it follows them where the PCI is higher, in the lower part of the landslide. The estimate is computed for every unmasked pixel of the area shown, and the display is as in Figure 11 . Pixels masked by NSPC (h = 0.30 h = 0.30 h = 0.30 , from the 79 acquisitions of the test area) are shown in gray. (a) Single-look interferogram. (b) Single-look interferogram of the first and last acquisitions of the substack. (c) PCI η \eta η , drawn on [ 0 , 0.6 ] [0, 0.6] [ 0 , 0.6 ] . (d) FaSHPS and (e) FaSHPS+ANS for the same 12-day interferogram, with a search window of 25 × 225 25 \times225 25 × 225 pixels. (f) Ensemble size of FaSHPS+ANS.
On the Maiun test area, where the PCI is low, ANS shrinks the search window little. The median η \eta η of the test area, 0.13, is close to that of the simulated class c = 0.35 c = 0.35 c = 0.35 with 79 acquisitions (Table 1 ), since η \eta η responds to the coherence of the short intervals. Few centers lie below the null quantile of S1-79, and at the others the PCI-scaled window is only somewhat smaller than the search window. Over all unmasked centers of the test area it holds a median of 3835 of the 4437 pixels of the search window. § 14.0.7 of the supplement gives further values of the phase-linked maps.
Runtime
With ANS the processing time of every SHP selection is lower than with the truncation step alone. Table 4 reports the processing time per center pixel on the real stacks, i.e., the time for the PCI-scaled window where it is used, the SHP selection, the truncation step where it is used, and the coherence matrix over the selected ensemble. Preparations that are performed once per image pixel, such as the PCI, are included. Times are relative to the full window of the same scene, because absolute times depend on the machine. Section S4.9 of the supplement describes the measurement.
Table 4. Processing time per center pixel up to the coherence matrix (§ 3.1.4 ) relative to the full window of the same scene, and mean ensemble size in pixels over the timed centers (in parentheses). Search windows of 8021 pixels (Visp) and 4437 pixels (Sentinel-1). Median over five interleaved passes. T marks the truncation step, and ANS the PCI-scaled window with the truncation step gated by η \eta η . The last row gives the time of the NSPC map of the mask per image pixel, relative to the full window per center pixel.
Without the truncation step, the processing with FaSHPS takes about half to three quarters of the time with the full window, since its test takes less time than its smaller ensembles save in the coherence matrix, and that with CMP a similar time. The truncation step adds up to half the processing time of the full window. The full window with the small search window of 31 × 31 pixels takes a tenth of the processing time with the 101 × 101 window on Visp (Table 4 ), but its map is a patchwork over vegetation (Figure 10 b). Relative to the processing with the SHP selection alone, ANS adds a little time on Visp and on the Maiun test area and saves time on the slow-deformation and long-term coherent test areas, where the smaller window and the truncated ensembles save more time than the truncation step takes. Either way the difference is within about 15 %. For every SHP selection and scene, the processing with ANS thus takes 14 to 32 % less time than with the truncation step alone, and that with FaSHPS+ANS less time than with the full window alone in every scene. On two other processors, an Intel Core i7-9700 and an AMD Ryzen 7 4800H, the processing with ANS also takes less time than with the truncation step alone. The NSPC map of the mask for the 1000 × 1000 pixels of the Visp image takes about as long as the full window at 75 center pixels, 0.35 s on one core of the test machine.
On quantitative comparisons with real data
We report no quantitative comparison between the SHP selections on real data, because neither of two measures without ground truth ranks them reliably in simulation, where the true phase is known
(Section S4.10 of the supplement). The split-stack agreement is the corrected SPC, over the even acquisitions, between the single-look phases of the center and the split-stack history, i.e., the phase history estimated from an ensemble that is selected on the odd acquisitions without the center and the neighbors that share its speckle. Between the full window and a pairwise SHP selection, both with ANS, it ranks as the phase linking error of the split-stack history does in only 42 to 56 % of the random evaluation scenes in which these errors differ by more than 5 % (TSX-44 and S1-79). It is also not independent of what it evaluates, since it is a corrected SPC with the center, the kind of agreement by which the truncation step chooses the ensemble. The predicted phase standard deviation, from the Cramér–Rao bound (Rodriguez and Martin, 1992, Eq. (25)) with the coherence and the effective number of looks of the ensemble, describes only the variance of an average, without its bias with respect to the center. It ranks the full window alone first in about three quarters of the random scenes, whereas the phase linking error rarely does so.
Between an SHP selection alone and the same SHP selection with ANS, the split-stack agreement ranks as the phase linking error of the split-stack history does in most of the scenes in which these errors differ by more than 5 %. On 16 regions of the real stacks (§ 14.0.10 of the supplement), it is higher with ANS in all 19 pairs of SHP selection and region whose values differ by more than 5 %. Since the measure prefers ANS in simulation somewhat more often than the phase linking error of the split-stack history does (§ 14.0.10 of the supplement), this is modest support from real data for ANS. Beyond this, the comparison of the SHP selections in this paper rests on the simulation study of § 2.3 , and the results on real data are shown as maps.
Discussion
Use of the mask and the PCI The mask and ANS can be applied separately, and ANS uses the PCI. The mask is computationally inexpensive and needs no data beyond the stack after topographic phase removal, and its threshold should be set relative to the correlated-noise floor of the stack (§ 2.2.1 ). Because it flags pixels in the shadow of buildings that a terrain model map misses and disagreement that neither map describes (§ 3.1.2 ), it complements geometric layover and shadow maps, also those derived from a surface model. The PCI rises with the local coherence and is available before any SHP selection. In this paper it sets the thresholds of the SHP tests, the scale of the search window and the gate, and it can serve wherever another processing step depends on the local coherence.
Use of ANS With ANS, a search window large enough for the decorrelated areas adds little error elsewhere (Figure 6 ), so that the user no longer has to trade the decorrelated areas against the fringes when choosing the window. Neither part of ANS compensates for a window that is too small. The truncation step of ANS can follow any SHP selection already in use. At one fringe or more per window diameter, most of the reduction by ANS comes from the truncation step (Table 3 ). The PCI-scaled window gives a coherent center fewer looks, since such a center reaches the same precision with fewer, and the gate skips the truncation step where the PCI shows no coherence. The processing with ANS is therefore faster than with the truncation step alone, at about the same error over all centers and with less error at one fringe or more per window diameter (§ 2.3.2 and § 3.1.4 ). Where clutter of moderate or high coherence is uniform over large parts of the window, both parts raise the error, ANS more than the truncation step alone, since some ensembles are truncated by chance and the PCI-scaled window gives up looks that need not be excluded (Table S9 of the supplement). In absolute terms the error added in the homogeneous elementary scenes with c = 0.5 c = 0.5 c = 0.5 and c = 0.95 c = 0.95 c = 0.95 is about a degree (TSX-44), against more than ten degrees that ANS removes at one fringe or more per window diameter in the random scenes. Where large uniform coherent areas dominate a scene, the truncation step alone is preferable to ANS, and where such areas need the highest precision, the full window alone.
SHP selections with ANS In the simulations, FaSHPS and CMP with ANS have slightly less error over all centers than the full window with ANS, and the choice of the search window changes the error more than the choice of SHP selection, except on point scatterers, where the choice of SHP selection matters more (§ 2.3.2 ). The simulated classes, however, differ in backscatter by at most 6 dB and contain no structures a few pixels wide, so that the simulations do not penalize a permissive significance level of the SHP tests. On Visp, the full window with ANS blends roads with the surrounding forest, whereas FaSHPS+ANS keeps them as lines of distinct phase (§ 3.1.3 , Figure 10 m to Figure 10 p). In scenes with such structures, the choice of SHP selection can thus matter more than in the simulations. On coherent targets, FaSHPS accepts most candidates of similar brightness, and neither SHP test separates them by phase, since the phase term of CMP on its own rejects practically no candidate (§ 2.2.5 ). The discrimination among such candidates is left to the truncation step, which truncates at a phase boundary only where the boundary lies close to the center, because the candidate ensembles are nested around it (Section S3.9 of the supplement).
Variants not adopted A variant of FaSHPS also accepts candidates that the amplitude test has rejected, if they lie outside the guard set and their SPC with the center exceeds the 99th percentile of S p q S_{pq} S pq in clutter with zero coherence. This threshold tests only against the absence of coherence. The variant therefore accepts coherent candidates on another part of a fringe or across a phase boundary, unless their phase difference to the center changes strongly over the acquisitions (Eq. (12) ). It gives FaSHPS+ANS slightly more error (Table S20 of the supplement). In place of the amplitude test of FaSHPS, the Kolmogorov–Smirnov, Kuiper, Anderson–Darling and generalized likelihood ratio tests have an error within 3 % of that of FaSHPS when their thresholds depend on the PCI, and 1.20 to 1.94 times its error with the thresholds for independent samples (TSX-44 and S1-79, Table S21 of the supplement). The calibration of the threshold thus matters more than the choice of the test. A local fringe frequency compensation before the truncation step hardly changed the error over all centers, and the local polynomial fit with ICI applied to every interferogram (Katkovnik et al., [8 ] ) had a higher error than FaSHPS+ANS (§ 14.1 of the supplement). Other variants of the truncation step were examined and not adopted. The first is the maximum of a Fisher information ∣ E ∣ Q ‾ / ( 1 − Q ‾ ) |\mathcal{E}|\overline{Q}/(1-\overline{Q}) ∣ E ∣ Q / ( 1 − Q ) over ensembles grown in order of SPC, with Q ‾ \overline{Q} Q from Eq. (6) . The others are ICI (Katkovnik, 1999) on a reference history, proxies of the mean squared error, a window derived from the second-moment matrix (Eq. (20) ), an order of all members by SPC, and a score computed from a full phase linking estimate in place of the multilook (Eq. (22) ). Of these, the rule of largest Fisher information and the order by SPC have more than 10 % more error over all centers than the same SHP selection with ANS (Table S20 of the supplement).
Table S21. Distribution tests in place of the amplitude test of FaSHPS, at a significance level of 1 %. Each entry is the phase linking error over all centers of the evaluation scenes relative to that of FaSHPS with the threshold that depends on η \eta η and the same variant (alone, with the truncation step, and with ANS), as a ratio of the weighted geometric means on the same centers. KS denotes the Kolmogorov–Smirnov test, AD the Anderson–Darling test and GLRT the generalized likelihood ratio test. The threshold that depends on η \eta η is calibrated in homogeneous clutter like that of FaSHPS (Section S2.2), and the threshold for independent samples is the nominal one. that depends on η \eta η , the error over all centers is 0.99 to 1.02 times that of FaSHPS with the same variant, and with the threshold for independent samples 1.11 to 1.94 times (alone, with the truncation step and with ANS, three configurations).
Limitations Several limitations remain. (i) The comparison of the methods rests on simulations with a CCG model. Neither of the two tested measures without ground truth ranks the SHP selections reliably in simulation. On real data the phase-linked results are therefore shown only as maps, and the split-stack agreement supports ANS only modestly (§ 3.1.5 ). (ii) The simulated phase consists of the deformation, which grows linearly in time, the atmosphere and, at point scatterers, the phase of their height above the ground (§ 2.3.1 ). Residual topographic phase of the clutter, e.g., from errors of the DEM, and closure phases, which arise, e.g., from changes of soil moisture (De Zan et al., 2014) and bias the estimated deformation (Ansari et al., 2021), are not simulated, so the comparison of the methods does not cover them. (iii) The score of the truncation step is not sensitive enough to truncate every ensemble that is too large, in particular at low coherence (Figure 5 ). With ANS, the phase linking error of FaSHPS exceeds twice the oracle error at 13 % (TSX-44) and 9.7 % (S1-79) of the area of the evaluation scenes, about a third of the share of FaSHPS alone, and a smaller marginal does not remove the ensembles that remain too large (Section S3.1 of the supplement). (iv) The thresholds of FaSHPS and CMP are set in simulated clutter that follows the exponential coherence model (Eq. (27) ), and where the coherence of a stack decays differently, their false rejection rate can depart from the significance level. (v) Point scatterers are not treated (§ 2.2.7 ). (vi) The NSPC and PCI maps and the other steps of § 2.2 are formulated for SLCs in radar coordinates, where the speckle correlation is the product of the correlations along the two image axes. (vii) Stacks with fewer than 20 acquisitions were not tested. (viii) The bound ρ min \rho_{\min} ρ m i n , below which a neighbor counts as free of shared speckle, is tuned once for all configurations. A neighbor at this bound keeps an expected corrected SPC of g ( ρ min ) 2 g(\rho_{\min})^2 g ( ρ m i n ) 2 , which does not fall with N N N , whereas the noise of one SPC value does, so that for stacks far longer than 79 acquisitions the bound would have to be lowered. It also trades distributed centers against point scatterers (Sections S3.1 and § 14.1 of the supplement). (ix) Of the published SHP selections, only FaSHPS and CMP are compared in full, CMP only with the single-reference interferograms of its authors (Section S2.1 of the supplement), and the distribution tests only in § 14.1 of the supplement. The tests on covariance matrices and the clustering by phase moments of § 2.2.5 are not compared.
Conclusions
We examined spatial phase coherence (SPC), a known pairwise quantity, as a basis for masking and for the size of the ensembles in phase estimation. Its expectation under the CCG model shows how decorrelation and differences in phase history each lower it. The speckle that adjacent pixels share sets a floor for it, which the processing parameters predict. Its mean over all pairs of an ensemble equals the mean corrected squared phase-only coherence of that ensemble. A threshold on its mean over the direct neighbors of each pixel gives a mask that covers a large part of the geometric layover on TerraSAR-X and Sentinel-1 data and, on Sentinel-1, a similar part of the shadow. On TerraSAR-X it also flags pixels in the shadow of buildings that a terrain model misses. On both sensors it flags further pixels that the layover and shadow maps used here do not explain. The mask thus complements geometrically derived layover and shadow maps. The mean of SPC over neighbors that share no speckle with the pixel, the PCI, rises with the local coherence and is available before any SHP selection, and it can assist the calibration of SHP tests that compare temporal mean amplitudes or intensities, such as those of FaSHPS and CMP. SPC requires no reference acquisition, no coherence model and no external data, and its behavior under noise, including speckle shared by neighbors, can be stated explicitly.
ANS, the PCI-scaled window with the gated truncation step, makes the size of the search window less critical, so that the window can be chosen large enough for the decorrelated areas. In simulations (TSX-44 and S1-79), over all centers, a window large enough for decorrelated areas has the lowest error of the tested sizes with ANS and 9.6 to 11 % less error than the same SHP selection without either step at its best size, for the full window, FaSHPS and CMP alike. On the real stacks, the processing with ANS up to the coherence matrix takes within about 15 % of the time with the SHP selection alone and 14 to 32 % less than with the truncation step alone, whose error ANS matches over all centers in the simulations. On Sentinel-1 data, ANS preserves the dense fringes that accumulate on a slowly deforming slope, where the SHP selections alone give broad patches separated by phase jumps. On a fast landslide, where dense fringes coincide with strong decorrelation, FaSHPS+ANS follows them only in the lower part of the landslide, where the PCI is higher. In large uniform coherent areas the error is small, and ANS raises it by about a degree (TSX-44). There the truncation step alone is preferable, and where the highest precision is needed, the full window without either step.
Several points remain open. A quantitative assessment on real data needs a measure without ground truth that ranks the SHP selections as the phase linking error does (§ 3.1.5 ). The score of the truncation step should become more sensitive at low coherence, where its noise lets ensembles that are too large pass. ANS has yet to be tested on stacks of fewer than 20 acquisitions, where the score rests on few interferograms. Point scatterers, which receive no separate treatment here, need their own handling. A PCI solved for each pixel as in Vecchioli et al. (2023) would give the coherence of the pixel itself instead of an average over its neighborhood. Stacks of geocoded SLCs, finally, need a model of the speckle correlation on the map grid, where it no longer separates along the image axes.
Abbreviations and notation
The lists below give the abbreviations and the mathematical notation of the paper, each symbol with the section or equation where it is defined.
Abbreviations
Mathematical notation
Table 5.
CRediT authorship contribution statement
Magnus Heimpel: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Irena Hajnsek: Supervision, Resources. Othmar Frey: Writing – review & editing, Validation, Supervision, Resources, Project administration, Methodology, Funding acquisition, Conceptualization.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Declaration of generative AI and AI-assisted technologies in the manuscript preparation process
During the preparation of this work the authors used Claude (Anthropic) in order to write and run simulation and analysis code, to analyze the simulation results, to produce figures and tables, to check the reference list, and to draft and revise the text of the manuscript. The research question, the method, the data and the decisions on what to test and report were the authors’. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.
Funding
This work was funded by the Swiss National Science Foundation (SNSF), https://data.snf.ch/grants/grant/216526 .
Data availability
The TerraSAR-X data of Visp used in this study, © DLR 2024, was acquired by the TerraSAR-X satellite of the German Aerospace Center (DLR) and provided by DLR via an announcement of opportunities (MTH3910). The acquisition dates are listed in Heimpel et al. (2026). The Sentinel-1 data is freely available from the Copernicus Data Space Ecosystem (https://dataspace.copernicus.eu ), and this paper contains modified Copernicus Sentinel data (2023 to 2025). The digital elevation models swissALTI3D and swissSURFACE3D Raster are freely available from swisstopo (source: Federal Office of Topography swisstopo), and the Copernicus DEM from the Copernicus Data Space Ecosystem. The topographic phase and the layover and shadow map of the Sentinel-1 test areas were produced using Copernicus WorldDEM-30 © DLR e.V. 2010-2014 and © Airbus Defence and Space GmbH 2014-2018 provided under COPERNICUS by the European Union and ESA; all rights reserved.
Code availability
The experiment code, with the mask, the PCI, the SHP selections, the PCI-scaled window, the truncation step and the scripts that produce every table and number of this paper, will be made available upon publication.
Acknowledgments
We would like to thank Marco Manzoni and Ehtasham Naseer for fruitful discussions, and the German Aerospace Center (DLR) for acquiring and providing the TerraSAR-X data.
Processing of a stack
Algorithm A.1 gives the processing of a stack with FaSHPS+ANS (Figure 2 ). With another SHP selection, lines 6 to 8 are replaced by that SHP selection. Three details do not appear in the algorithm. The PCI (§ 2.2.2 ) is computed from all pixels, masked ones included, except that a pixel with a zero value in some acquisition (0.9 % of Visp) enters the PCI of no neighbor. Masked pixels are not part of the core of the truncation step, and the second-moment matrix that orders the members of an ensemble (Eq. (20) ) leaves out masked pixels and pixels with a zero value in some acquisition.
Algorithm A.1 Processing of a stack with FaSHPS+ANS
Require: SLC stack, simulated topographic phase, search window B \mathcal{B} B , ring R p \mathcal{R}_p R p (§ 2.2.2 ), core C p \mathcal{C}_p C p and guard set G p \mathcal{G}_p G p (§ 2.2.6 ) from the predicted speckle correlation, mask threshold h h h , thresholds z 1 z_1 z 1 and z α ( η ) z_\alpha(\eta) z α ( η ) (§ 2.2.5 ), null quantile η 0 \eta_0 η 0 , settings of Table 2
Per stack
V ← \mathcal{V} \leftarrow V ← the pixels with finite, nonzero observations in all acquisitions
subtract the simulated topographic phase, form unit phasors u q u_q u q and temporal mean amplitudes A ‾ q \overline{A}_q A q
NSPC (Eq. (14) ) for p ∈ V p \in\mathcal{V} p ∈ V and the mask V h \mathcal{V}_h V h
η ~ \widetilde{\eta} η (Eq. (15) ) and its 5 × 5 5 \times5 5 × 5 average η \eta η for every pixel
Per center p ∈ V h p \in\mathcal{V}_h p ∈ V h
B p ′ ← \mathcal{B}'_p \leftarrow B p ′ ← the PCI-scaled window (Eq. (17) ), restricted to V h \mathcal{V}_h V h
form the reference set of p p p from p p p and the q ∈ B p ′ q \in\mathcal{B}'_p q ∈ B p ′ with T p q ≤ z 1 T_{pq} \leq z_1 T pq ≤ z 1 for μ p = A ‾ p \mu_p = \overline{A}_p μ p = A p (Eq. (19) ), then set μ p \mu_p μ p to the mean of A ‾ q \overline{A}_q A q over it
accept q ∈ B p ′ q \in\mathcal{B}'_p q ∈ B p ′ if the second-pass statistic T p q T_{pq} T pq is at most z α ( η p ) z_\alpha(\eta_p) z α ( η p )
O p ← \mathcal{O}_p \leftarrow O p ← the eight-connected component of the accepted pixels that contains p p p
if ∣ O p ∣ < L min \lvert\mathcal{O}_p\rvert< L_{\min} ∣ O p ∣ < L m i n then
p p p is insufficient and keeps its single-look data
else
E ← O p \mathcal{E} \leftarrow\mathcal{O}_p E ← O p
if η p ≥ η 0 \eta_p \geq\eta_0 η p ≥ η 0 and at least ⌈ 1.5 L min ⌉ = 38 \lceil1.5L_{\min}\rceil= 38 ⌈ 1.5 L m i n ⌉ = 38 members of O p \mathcal{O}_p O p lie outside G p \mathcal{G}_p G p (Section S2.3) then
sort O p \mathcal{O}_p O p in the ellipse order (Eq. (20) ), or in the distance order where it is not defined (Section S2.3 of the supplement)
form the running sums of u q ( n ) u q ( r ) ‾ u_q(n)\overline{u_q(r)} u q ( n ) u q ( r ) over the sorted members outside G p \mathcal{G}_p G p , for ( n , r ) ∈ I (n,r) \in\mathcal{I} ( n , r ) ∈ I (§ 2.2.6 and § 2.2.9 )
for the sizes L 1 = 25 < L 2 < ⋯ < L K L_1 = 25 < L_2 < \cdots< L_K L 1 = 25 < L 2 < ⋯ < L K of Eq. (21) (Section S2.3 of the supplement) do
a k ← a_k \leftarrow a k ← score (Eq. (23) ) of E k \mathcal{E}_k E k , and σ k \sigma_k σ k (Eq. (25) )
end for
k max ← arg max k a k k_{\max} \leftarrow\arg\max_k a_k k m a x ← arg max k a k , σ Δ a \sigma_{\Delta a} σ Δ a by Eq. (39) , ζ ← κ σ Δ a \zeta\leftarrow\kappa\sigma_{\Delta a} ζ ← κ σ Δ a
if a k max − a K > ζ a_{k_{\max}} - a_K > \zeta a k m a x − a K > ζ and a k max > λ σ k max a_{k_{\max}} > \lambda\sigma_{k_{\max}} a k m a x > λ σ k m a x then
k ⋆ ← max { k : a k ≥ a k max − ζ / 3 } k^\star\leftarrow\max\{k : a_k \geq a_{k_{\max}} - \zeta/3\} k ⋆ ← max { k : a k ≥ a k m a x − ζ /3 } , E ← \mathcal{E} \leftarrow E ← the sorted members up to the L k ⋆ L_{k^\star} L k ⋆ -th scored one
end if
end if
Γ ^ ← \widehat{\Gamma} \leftarrow Γ ← Eq. (26) over E \mathcal{E} E , θ ^ ← \widehat{\theta} \leftarrow θ ← the maximizer of Eq. (28)
end if
supplement.pdf
This supplement holds the proofs, further tables, detailed analyses and options that complete the paper. References to sections, equations, tables and figures without the prefix S are to the paper. TSX-44, S1-79 and S1-20 are the simulated configurations of § 2.3.1 , named by sensor and number of acquisitions.
Spatial phase coherence
This section completes § 2.1 , § 2.2.1 and the estimate of the speckle correlation in § 3.1.1 .
Relation to existing pairwise phase measures
This section gives the details of the relations of § 2.1.2 .
Zero-lag phase cross-correlation At zero lag, the cross-correlation criterion of Narayan et al. [13 ] , Eq. (6) is ∑ i = 1 N I w p ( i ) w q ( i ) ‾ \sum_{i=1}^{N_I} w_p(i)\overline{w_q(i)} ∑ i = 1 N I w p ( i ) w q ( i ) with w p ( i ) = e i χ p ( i ) w_p(i)=e^{\mathrm{i}\chi_p(i)} w p ( i ) = e i χ p ( i ) the observed phasor of interferogram i i i , where N I N_I N I is the number of interferograms. Its magnitude is N I S p q N_I\sqrt{S_{pq}} N I S pq evaluated on the same vectors. Their rule compares correlations at different lags, whereas a threshold on SPC, the Rayleigh test of § 2.1.2 , compares the zero-lag magnitude with a fixed value. A maximum at zero lag does not by itself establish significance.
Centered complex correlation For vectors of unit phasors, write a p q = u p H u q / N a_{pq}=u_p^{H}u_q/N a pq = u p H u q / N and m p = ∑ n u p ( n ) / N m_p=\sum_n u_p(n)/N m p = ∑ n u p ( n ) / N . The centered correlation coefficient of Shi et al. [17 ] , Eq. (1) applied to these vectors is
r p q ( u ) = a p q − m p ‾ m q ( 1 − ∣ m p ∣ 2 ) ( 1 − ∣ m q ∣ 2 ) , S p q = ∣ a p q ∣ 2 , (32) r_{pq}^{(u)}=\frac{a_{pq}-\overline{m_p}m_q}{\sqrt{(1-\lvert m_p\rvert^2)(1-\lvert m_q\rvert^2)}},\qquad S_{pq}=\lvert a_{pq}\rvert^2, \tag*{(32)} r pq ( u ) = ( 1 − ∣ m p ∣ 2 ) ( 1 − ∣ m q ∣ 2 ) a pq − m p m q , S pq = ∣ a pq ∣ 2 , (32)
provided both centered vectors are nonzero. The two coincide when m p = m q = 0 m_p=m_q=0 m p = m q = 0 but differ in general. For a temporally constant phase the centered coefficient is undefined, whereas SPC equals one. A constant rotation of either pixel leaves SPC and ∣ r p q ∣ \lvert r_{pq}\rvert ∣ r pq ∣ unchanged but changes arg r p q \arg r_{pq} arg r pq , so that a constraint on arg r p q \arg r_{pq} arg r pq distinguishes offsets that SPC treats as equivalent. An acquisition-dependent rotation common to both pixels leaves SPC unchanged but generally changes r p q r_{pq} r pq through the temporal means.
Circular variance of high-pass phases For each interferogram k k k , Pepe et al. [15 ] form the high-pass phasor exp ( i ( φ k ( q ) − φ k L P ) ) \exp\left(i(\varphi_k(q)-\varphi_k^{\mathrm{LP}})\right) exp ( i ( φ k ( q ) − φ k LP ) ) of pixel q q q , where φ k L P \varphi_k^{\mathrm{LP}} φ k LP is the phase of the mean of the unit phasors over the window (their Eqs. (2) to (4)). The mean resultant length of these phasors over the interferograms (their Eq. (9)) is S \sqrt{S} S between the phasor vector of q q q and the unit phasors of the low-pass phases, and the circular variance (their Eq. (10)) is one minus this value. Pixels with a circular variance below 0.7 are retained, and each of them is compared with the retained pixel of lowest circular variance by Kuiper’s test (Kuiper [9 ] ) at a significance level of 20 %.
Pairwise phase consistency and phase-locking value Vinck et al. [21 ] define the pairwise phase consistency as the mean over pairs of trials of the cosine of the difference of relative phases, which is Eq. (5) with trials in place of acquisitions. They show (their Eqs. (13) and (19)) that its expected value equals the square of the population value of the phase-locking value of Lachaux et al. [10 ] and that, unlike the squared sample phase-locking value, its sample estimate is free of finite-sample bias.
Proofs
This section gives the proofs of Eqs. (5) , (6) , (8) to (10) and (13) of § 2.1.3 to § 2.1.6 , of the bound on the expected corrected SPC of direct neighbors in § 2.2.1 , and of the standard deviation of corrected SPC under independent uniform phases that Section S3.1 uses. The heading of each paragraph names what it proves.
Agreement identity, (5) For z n = u p ( n ) u q ( n ) z_n=u_p(n)u_q(n) z n = u p ( n ) u q ( n ) ,
N 2 S p q = N + 2 ∑ s < t Re ( z t ‾ z s ) = N + 2 ∑ s < t cos ( Δ ϕ p s t − Δ ϕ q s t ) , (33) N^2S_{pq}=N+2\sum_{s<t}\operatorname{Re}(\overline{z_t}z_s)=N+2\sum_{s<t}\cos(\Delta\phi_p^{st}-\Delta\phi_q^{st}), \tag*{(33)} N 2 S pq = N + 2 s < t ∑ Re ( z t z s ) = N + 2 s < t ∑ cos ( Δ ϕ p s t − Δ ϕ q s t ) , (33)
and rearranging gives Eq. (5) .
Coherence of an ensemble, (6) With v q s t = u q ( s ) u q ( t ) ‾ v_q^{st}=u_q(s)\overline{u_q(t)} v q s t = u q ( s ) u q ( t ) ,
γ ^ s t ∘ ( E ) 2 = 1 L 2 ∑ q , q ′ ∈ E v q s t v q ′ s t ‾ = 1 L + 1 L 2 ∑ q ≠ q ′ cos ( Δ ϕ q s t − Δ ϕ q ′ s t ) , (34) \widehat{\gamma}_{st}^{\circ}(\mathcal{E})^2=\frac{1}{L^2}\sum_{q,q'\in\mathcal{E}}v_q^{st}\overline{v_{q'}^{st}}=\frac{1}{L}+\frac{1}{L^2}\sum_{q\ne q'}\cos(\Delta\phi_q^{st}-\Delta\phi_{q'}^{st}), \tag*{(34)} γ s t ∘ ( E ) 2 = L 2 1 q , q ′ ∈ E ∑ v q s t v q ′ s t = L 1 + L 2 1 q = q ′ ∑ cos ( Δ ϕ q s t − Δ ϕ q ′ s t ) , (34)
since the off-diagonal terms occur in conjugate pairs. Averaging over s < t s<t s < t and applying Eq. (5) to each pair gives
1 M ∑ s < t γ ^ s t ∘ ( E ) 2 = 1 L + L − 1 L Q ‾ ( E ) , \frac{1}{M}\sum_{s<t}\widehat{\gamma}_{st}^{\circ}(\mathcal{E})^2=\frac{1}{L}+\frac{L-1}{L}\overline{Q}(\mathcal{E}), M 1 s < t ∑ γ s t ∘ ( E ) 2 = L 1 + L L − 1 Q ( E ) ,
which rearranges to Eq. (6) .
Expectations, Eqs. (8) to (10) and (13) Each cosine in Eq. (5) is cos ( Δ ϕ p s t − Δ ϕ q s t ) = Re v p s t v q s t ‾ \cos(\Delta\phi_p^{st}-\Delta\phi_q^{st})=\operatorname{Re}\,v_p^{st}\overline{v_q^{st}} cos ( Δ ϕ p s t − Δ ϕ q s t ) = Re v p s t v q s t , and for independent pixels its expectation factorizes as Re E [ v p s t ] E [ v q s t ‾ ] \operatorname{Re}\,\mathbb{E}[v_p^{st}]\mathbb{E}[\overline{v_q^{st}}] Re E [ v p s t ] E [ v q s t ] . For a CCG pixel, write Δ ϕ p s t = Δ θ p s t + ϵ \Delta\phi_p^{st}=\Delta\theta_p^{st}+\epsilon Δ ϕ p s t = Δ θ p s t + ϵ with the phase error ϵ = Δ ϕ p s t − arg γ s t p \epsilon=\Delta\phi_p^{st}-\arg\gamma_{st}^{p} ϵ = Δ ϕ p s t − arg γ s t p , so that
E [ v p s t ] = e i Δ θ p s t ( E [ cos ϵ ] + i E [ sin ϵ ] ) . \mathbb{E}[v_p^{st}]=e^{i\Delta\theta_p^{st}}\left(\mathbb{E}[\cos\epsilon]+i\mathbb{E}[\sin\epsilon]\right). E [ v p s t ] = e i Δ θ p s t ( E [ cos ϵ ] + i E [ sin ϵ ] ) .
The acquisitions s s s and t t t of a CCG vector form a bivariate CCG pair with correlation γ s t p \gamma_{st}^{p} γ s t p , so that E [ cos ϵ ] = g ( ∣ γ s t p ∣ ) \mathbb{E}[\cos\epsilon]=g(\lvert\gamma_{st}^{p}\rvert) E [ cos ϵ ] = g (∣ γ s t p ∣) by Eq. (7) . After the second acquisition is multiplied by e i arg γ s t p e^{i\arg\gamma_{st}^{p}} e i a r g γ s t p , the pair has the real correlation ∣ γ s t p ∣ \lvert\gamma_{st}^{p}\rvert ∣ γ s t p ∣ and a real covariance matrix, so that its distribution is invariant under complex conjugation. Since conjugation maps ϵ \epsilon ϵ to − ϵ -\epsilon − ϵ , the distribution of ϵ \epsilon ϵ is symmetric about zero, E [ sin ϵ ] = 0 \mathbb{E}[\sin\epsilon]=0 E [ sin ϵ ] = 0 , and E [ v p s t ] = g ( ∣ γ s t p ∣ ) e i Δ θ p s t \operatorname{E}[v_p^{st}] = g(|\gamma_{st}^{p}|)e^{i\Delta\theta_{p}^{st}} E [ v p s t ] = g ( ∣ γ s t p ∣ ) e i Δ θ p s t [12 , 11 Eq. (21] (Eqs. (8) and (9); cf. Just and Bamler 1994). With this for both pixels, the expectation of each cosine is g ( ∣ γ s t p ∣ ) g ( ∣ γ s t q ∣ ) cos ( Δ θ p s t − Δ θ q s t ) g(|\gamma_{st}^{p}|)g(|\gamma_{st}^{q}|)\cos(\Delta\theta_{p}^{st}-\Delta\theta_{q}^{st}) g ( ∣ γ s t p ∣ ) g ( ∣ γ s t q ∣ ) cos ( Δ θ p s t − Δ θ q s t ) , and the mean over s < t s<t s < t gives Eq. (8) . For relative phasors independent across acquisitions,
E [ ∣ ∑ n z n ∣ 2 ] = N + ∑ n ≠ m ξ n ξ m ‾ = N + ∣ ∑ n ξ n ∣ 2 − ∑ n ∣ ξ n ∣ 2 , \operatorname{E}\left[\left|\sum_n z_n\right|^2\right] = N + \sum_{n\ne m}\xi_n\overline{\xi_m} = N + \left|\sum_n \xi_n\right|^2 - \sum_n|\xi_n|^2, E n ∑ z n 2 = N + n = m ∑ ξ n ξ m = N + n ∑ ξ n 2 − n ∑ ∣ ξ n ∣ 2 ,
where the N N N comes from ∣ z n ∣ = 1 |z_n|=1 ∣ z n ∣ = 1 and the sum over n ≠ m n\ne m n = m from the independence, and Eq. (9) follows from Eq. (4) . For a bivariate CCG pair with correlation ρ \rho ρ , E [ z n ] = g ( ∣ ρ ∣ ) e i arg ρ \operatorname{E}[z_n]=g(|\rho|)e^{i\arg\rho} E [ z n ] = g ( ∣ ρ ∣ ) e i a r g ρ , which gives Eq. (10) . Multiplying each z n z_n z n by the deterministic phasor e − i ψ p q ( n ) e^{-i\psi_{pq}(n)} e − i ψ pq ( n ) gives Eq. (13) .
Floor under temporal coherence, Section 3.1. Eq. (8) assumes independent pixels and Eq. (9) independent acquisitions. Direct neighbors with a coherence between zero and one satisfy neither: their values are correlated between the two pixels through the shared speckle and between acquisitions through the coherence. The floor of Eq. (10) remains a lower bound under the separable model of the simulations, in which the covariance of two pixel values is the product of the temporal coherence of their acquisitions and the speckle correlation of the two pixels. Let the pixel values x ( n ) = ( x p ( n ) , x q ( n ) ) x(n)=(x_p(n),x_q(n)) x ( n ) = ( x p ( n ) , x q ( n )) be zero-mean CCG with Cov [ x a ( s ) , x b ( t ) ] = γ s t ρ a b \operatorname{Cov}[x_a(s),x_b(t)]=\gamma_{st}\rho_{ab} Cov [ x a ( s ) , x b ( t )] = γ s t ρ ab for a , b ∈ { p , q } a,b\in\{p,q\} a , b ∈ { p , q } , where ρ p p = ρ q q = 1 \rho_{pp}=\rho_{qq}=1 ρ pp = ρ q q = 1 , ρ p q = ρ \rho_{pq}=\rho ρ pq = ρ , γ s s = 1 \gamma_{ss}=1 γ ss = 1 and 0 ≤ γ s t ≤ 1 0\le\gamma_{st}\le1 0 ≤ γ s t ≤ 1 once a true phase history common to both pixels is removed, which leaves every z n z_n z n unchanged.
By Eq. (5) , since cos ( Δ ϕ p s t − Δ ϕ q s t ) = Re ( z t z s ‾ ) \cos(\Delta\phi_p^{st}-\Delta\phi_q^{st})=\operatorname{Re}(z_t\overline{z_s}) cos ( Δ ϕ p s t − Δ ϕ q s t ) = Re ( z t z s ) ,
E [ Q p q ] = 1 M ∑ s < t Re E [ z t z s ‾ ] , \operatorname{E}[Q_{pq}] = \frac{1}{M}\sum_{s<t}\operatorname{Re}\operatorname{E}[z_t\overline{z_s}], E [ Q pq ] = M 1 s < t ∑ Re E [ z t z s ] ,
and each term involves only the acquisitions s s s and t t t . For one such pair with γ = γ s t \gamma=\gamma_{st} γ = γ s t , let a = ( a p , a q ) a=(a_p,a_q) a = ( a p , a q ) , b s = ( b s , p , b s , q ) b_s=(b_{s,p},b_{s,q}) b s = ( b s , p , b s , q ) and b t = ( b t , p , b t , q ) b_t=(b_{t,p},b_{t,q}) b t = ( b t , p , b t , q ) be independent pairs of zero-mean CCG values with unit variance and correlation ρ \rho ρ , and
x c ( s ) = γ a c + 1 − γ b s , c , x c ( t ) = γ a c + 1 − γ b t , c , c ∈ { p , q } . x_c(s)=\sqrt{\gamma}\,a_c+\sqrt{1-\gamma}\,b_{s,c},\qquad x_c(t)=\sqrt{\gamma}\,a_c+\sqrt{1-\gamma}\,b_{t,c},\qquad c\in\{p,q\}. x c ( s ) = γ a c + 1 − γ b s , c , x c ( t ) = γ a c + 1 − γ b t , c , c ∈ { p , q } .
The pair a a a is common to both acquisitions, and b s b_s b s and b t b_t b t are renewed in each. These values have unit variance, correlation ρ \rho ρ between the two pixels in each acquisition, γ \gamma γ between the two acquisitions of one pixel and γ ρ \gamma\rho γ ρ between different pixels in different acquisitions. They thus have the covariances of the model and, as zero-mean CCG values, its distribution.
The phasor z s z_s z s depends only on x ( s ) x(s) x ( s ) , hence on a a a and b s b_s b s , and z t z_t z t only on a a a and b t b_t b t . Given a a a , the two phasors are therefore independent, and they have the same conditional mean m ( a ) = E [ z s ∣ a ] = E [ z t ∣ a ] m(a)=\operatorname{E}[z_s\mid a]=\operatorname{E}[z_t\mid a] m ( a ) = E [ z s ∣ a ] = E [ z t ∣ a ] , since b s b_s b s and b t b_t b t have the same distribution. Hence
E [ z t z s ‾ ] = E a [ E [ z t ∣ a ] E [ z s ‾ ∣ a ] ] = E a ∣ m ( a ) ∣ 2 = ∣ E a m ( a ) ∣ 2 + E a ∣ m ( a ) − E a m ( a ) ∣ 2 ≥ ∣ E [ z s ] ∣ 2 , \operatorname{E}[z_t\overline{z_s}] = \operatorname{E}_a\left[\operatorname{E}[z_t\mid a]\operatorname{E}[\overline{z_s}\mid a]\right] = \operatorname{E}_a|m(a)|^2 = |\operatorname{E}_a m(a)|^2+\operatorname{E}_a|m(a)-\operatorname{E}_a m(a)|^2 \ge|\operatorname{E}[z_s]|^2, E [ z t z s ] = E a [ E [ z t ∣ a ] E [ z s ∣ a ] ] = E a ∣ m ( a ) ∣ 2 = ∣ E a m ( a ) ∣ 2 + E a ∣ m ( a ) − E a m ( a ) ∣ 2 ≥ ∣ E [ z s ] ∣ 2 ,
where the third equality splits the mean square of m ( a ) m(a) m ( a ) into its squared mean and its variance. Alone, the pair x ( s ) x(s) x ( s ) is CCG with correlation ρ \rho ρ , so that E [ z s ] = g ( ∣ ρ ∣ ) e i arg ρ \operatorname{E}[z_s]=g(|\rho|)e^{i\arg\rho} E [ z s ] = g ( ∣ ρ ∣ ) e i a r g ρ as in the proof of Eq. (10) , and ∣ E [ z s ] ∣ 2 = g ( ∣ ρ ∣ ) 2 |\operatorname{E}[z_s]|^2=g(|\rho|)^2 ∣ E [ z s ] ∣ 2 = g ( ∣ ρ ∣ ) 2 for every pair of acquisitions. The mean over the pairs gives
E [ Q p q ] ≥ g ( ∣ ρ ∣ ) 2 . (35) \operatorname{E}[Q_{pq}] \ge g(|\rho|)^2. \tag*{(35)} E [ Q pq ] ≥ g ( ∣ ρ ∣ ) 2 . (35)
If γ s t = 0 \gamma_{st}=0 γ s t = 0 for all s ≠ t s\ne t s = t , the common part a a a drops out, m ( a ) m(a) m ( a ) is constant and the bound holds with equality, which is Eq. (10) . Temporal coherence can thus only raise the expected corrected SPC of direct neighbors above the floor. Signals that do not share the speckle correlation of the clutter, such as thermal noise or the sidelobes of a bright scatterer, are outside this model.
Null moments, Eq. (36) . Under independent uniform phases, the terms of E 0 ∣ ∑ n z n ∣ 4 \mathbb{E}_0\left|\sum_n z_n\right|^4 E 0 ∣ ∑ n z n ∣ 4 that pair each z n z_n z n with a z ˉ m \bar{z}_m z ˉ m are the only ones with nonzero mean, which gives
E 0 ∣ ∑ n z n ∣ 4 = 2 N 2 − N , Var 0 ( S p q ) = N − 1 N 3 , \mathbb{E}_0\left|\sum_n z_n\right|^4 = 2N^2-N,\qquad\operatorname{Var}_0(S_{pq}) = \frac{N-1}{N^3}, E 0 n ∑ z n 4 = 2 N 2 − N , Var 0 ( S pq ) = N 3 N − 1 ,
and, by Eq. (4) , the standard deviation of one corrected SPC value
sd 0 ( Q p q ) = 1 N ( N − 1 ) . (36) \operatorname{sd}_0(Q_{pq}) = \frac{1}{\sqrt{N(N-1)}}. \tag*{(36)} sd 0 ( Q pq ) = N ( N − 1 ) 1 . (36)
Aggregation of neighborhood SPC
This section completes the aggregation of SPC over the neighbors of a pixel in § 2.2.1 .
For m = ∣ N p ∣ > 0 m = |\mathcal{N}_p| > 0 m = ∣ N p ∣ > 0 valid neighbors, Eq. (14) is the first of
ℓ p m e a n = 1 m ∑ q Q p q + , ℓ p r m s = max ( 1 m ∑ q Q p q , 0 ) , \ell_p^{\mathrm{mean}} = \frac{1}{m}\sum_q \sqrt{Q_{pq}^{+}},\qquad\ell_p^{\mathrm{rms}} = \sqrt{\max\left(\frac{1}{m}\sum_q Q_{pq},0\right)}, ℓ p mean = m 1 q ∑ Q pq + , ℓ p rms = max ( m 1 q ∑ Q pq , 0 ) ,
and the second reverses averaging and square root. Let β p q = ∣ E [ u p ( n ) u q ( n ) ] ∣ \beta_{pq} = \left|\mathbb{E}[u_p(n)u_q(n)]\right| β pq = ∣ E [ u p ( n ) u q ( n )] ∣ under independent, identically distributed sampling. As N → ∞ N \to\infty N → ∞ , ℓ p m e a n → m − 1 ∑ q β p q \ell_p^{\mathrm{mean}} \to m^{-1}\sum_q \beta_{pq} ℓ p mean → m − 1 ∑ q β pq and ℓ p r m s → ( m − 1 ∑ q β p q 2 ) 1 / 2 \ell_p^{\mathrm{rms}} \to\left(m^{-1}\sum_q \beta_{pq}^2\right)^{1/2} ℓ p rms → ( m − 1 ∑ q β pq 2 ) 1/2 . For eight neighbors, one with β = 1 \beta= 1 β = 1 and seven with β = 0 \beta= 0 β = 0 , the limits of the mean and the RMS form are 0.125 and 0.354, respectively. For eight with β = 0.2 \beta= 0.2 β = 0.2 , both are 0.2. The RMS form favors a few strongly agreeing neighbors, whereas the mean form favors broad agreement, which is the intended criterion for the NSPC mask. At finite N N N , the mean form clips each term separately, so that either form can exceed the other. If all Q p q ≥ 0 Q_{pq} \ge0 Q pq ≥ 0 , ℓ p r m s ≥ ℓ p m e a n \ell_p^{\mathrm{rms}} \ge\ell_p^{\mathrm{mean}} ℓ p rms ≥ ℓ p mean .
Pixels that enter the estimated speckle correlation
The estimate (Eq. (31) ) of § 3.1.1 uses the pairs at lag one of the pixels with η < η 0 \eta< \eta_0 η < η 0 , so that the pairs share no signal beyond the speckle. Two further effects bias Q ˉ 1 \bar{Q}_1 Q ˉ 1 . First, a subtracted model phase such as the topographic phase attenuates Q p q Q_{pq} Q pq where it differs between neighbors in a way that changes over the acquisitions (§ 2.1.6 ). Only pairs whose model phase difference alone has a corrected SPC Q [ e i ψ p q ] Q[e^{i\psi_{pq}}] Q [ e i ψ pq ] of at least 0.9 (Eq. (13) ) enter Q ˉ 1 \bar{Q}_1 Q ˉ 1 , so that the subtraction lowers their expected corrected SPC by at most 10 %. Second, the correlation depends on the brightness of the surface. Thermal noise lowers it on dark surfaces, and point scatterers raise it on the brightest. Only pixels whose surroundings, a 9 × 9 9 \times9 9 × 9 window without its central 3 × 3 3 \times3 3 × 3 pixels, have a mean amplitude between the median and the 90th percentile of this quantity over the valid pixels enter. The mean amplitude of the surroundings is used rather than that of the pixel itself, because a selection by the pixel’s own amplitude favors pixels whose speckle happens to be bright in many acquisitions, and bright samples agree better in phase. In simulated clutter with N = 44 N = 44 N = 44 , such a selection raises the estimate by 0.03, whereas the selection by the surroundings leaves it unchanged. To extend the estimate to the other lags, the ratio B / f s B/f_s B / f s of a raised-cosine weighting with the weighting coefficient of the sensor is fitted to the lag-one value of each axis, and Eq. (11) gives the correlation at these lags.
On the Vısp stack, 61 % of the valid pixels have a PCI below its null quantile, and the pairs that enter the estimate are selected among them. Over all pairs of valid pixels of this scene, the mean corrected SPC at lag one is 0.33 in azimuth and 0.28 in range (Figure 7 a), close to the floor g ( ∣ ρ ∣ ) 2 g(|\rho|)^2 g ( ∣ ρ ∣ ) 2 of Eq. (10) for the predicted correlation (Table S11 ), as expected where most pixels decorrelate over the stack. On Sentinel-1, the estimate lies below the prediction in azimuth (§ 3.1.1 ). To find the cause, we use the correlation of the intensities of neighboring pixels over time, which measures ∣ ρ ∣ 2 |\rho|^2 ∣ ρ ∣ 2 from the intensities alone. In the five test areas, its square root at lag one is close to the prediction over bright surfaces. Over dark surfaces, where the mean amplitude of the surroundings lies more than 9 dB below the median of the test area, it decreases in azimuth to 0.43 to 0.46, as expected for thermal noise with a flat spectrum over the processed bandwidth. In simulated clutter with zero coherence and the predicted correlations, NSPC has a median of 0.54 (S1-79), whereas pure thermal noise over 79 acquisitions, with the azimuth correlation measured on dark surfaces or expected for a flat noise spectrum, has a median NSPC of 0.38 to 0.42.
SHP selection and adaptive sizing
This section completes Section 3.
Test of CMP
CMP ([24 ] ) compares a center p p p and a candidate q q q through their temporal mean intensities I ˉ p \bar{I}_p I ˉ p and I ˉ q \bar{I}_q I ˉ q and the phases Δ ϕ p 1 n \Delta\phi_p^{1n} Δ ϕ p 1 n and Δ ϕ q 1 n \Delta\phi_q^{1n} Δ ϕ q 1 n of the N − 1 N-1 N − 1 single-reference interferograms, whose reference is the first acquisition. Zhao et al. [24 ] note that the interferograms can follow a single-reference or a multi-reference scheme and use single-reference interferograms, with the first acquisition as reference in their simulations (their § 2.2.1 ). We follow this choice. A single reference makes every term of the statistic share the noise of that acquisition, and the first acquisition gives the interferograms the longest temporal baselines, which under the decaying coherence of Eq. (27) lowers their mean coherence. A central reference acquisition, several reference acquisitions or the interferograms of consecutive acquisitions may therefore serve CMP better, and they are not tested here. The statistic of CMP is the mean over these interferograms of a symmetrized Kullback–Leibler divergence between the 2 × 2 2 \times2 2 × 2 covariance matrices of the two pixels. Each matrix is built from the mean intensity of the pixel, its interferometric phase and an assumed coherence γ a ( n ) \gamma_a(n) γ a ( n ) of interferogram n n n , and the statistic reads (Zhao et al. [24 ] , Eqs. (1) , (3) to (5) and (10) )
D p q = 4 π I ˉ p 2 + I ˉ q 2 I ˉ p I ˉ q ( 1 N − 1 ∑ n = 2 N 1 − γ a 2 ( n ) cos ( Δ ϕ p 1 n − Δ ϕ q 1 n ) 1 − γ a 2 ( n ) ) − 8 π . (37) D_{pq} = \frac{4}{\pi}\frac{\bar{I}_p^2+\bar{I}_q^2}{\bar{I}_p\bar{I}_q}\left(\frac{1}{N-1}\sum_{n=2}^{N}\frac{1-\gamma_a^2(n)\cos\left(\Delta\phi_p^{1n}-\Delta\phi_q^{1n}\right)}{1-\gamma_a^2(n)}\right)-\frac{8}{\pi}. \tag*{(37)} D pq = π 4 I ˉ p I ˉ q I ˉ p 2 + I ˉ q 2 ( N − 1 1 n = 2 ∑ N 1 − γ a 2 ( n ) 1 − γ a 2 ( n ) cos ( Δ ϕ p 1 n − Δ ϕ q 1 n ) ) − π 8 . (37)
It is zero for equal intensities and phases and grows as the intensity ratio departs from one and with the phase differences. The assumed coherence γ a ( n ) \gamma_a(n) γ a ( n ) is the product of three factors (Zhao et al. [24 ] , Eqs. (8) and (9) and § 2.2.1 ). The first is the scale 0.3, which Zhao et al. [24 ] set for single-look data. The second is a thermal term 1 / ( 1 + S N R − 1 ) 1/(1+\mathrm{SNR}^{-1}) 1/ ( 1 + SNR − 1 ) of 0.94 for a signal-to-noise ratio SNR of 12 dB (Zebker and Villasenor, [23 ] , Eq. (4) ), against the 0.92 that Zhao et al. [24 ] state. The third is an exponential decay with the temporal baseline of the interferogram. We leave out their spatial term, which depends on the perpendicular baseline, and fit the decay constant, which Zhao et al. [24 ] set to 200 days, to the decay of the mean coherence of the single-reference interferograms with the temporal baseline, with the same estimator on simulated and real stacks (Sections S3.1 and § 14.0.1 ).
A candidate passes if D p q D_{pq} D pq is at most the threshold, which is the test of Zhao et al. [24 ] on exp ( − D p q ) \exp(-D_{pq}) exp ( − D pq ) . The test has two passes (Zhao et al. [24 ] ). The first compares every candidate with the center by Eq. (37) . The center and the candidates accepted in the first pass form the reference set. The second pass tests every candidate again, with the mean intensity of the reference set in place of I ˉ p \bar{I}_p I ˉ p and the phases of the multilooked interferograms of the reference set in place of Δ ϕ p 1 n \Delta\phi_p^{1n} Δ ϕ p 1 n . The ensemble is the eight-connected component of the center among the candidates that pass the second pass. Zhao et al. [24 ] use one threshold for both passes, which they set from simulated data at a significance level of 5% and which depends mainly on the stack size (Zhao et al., [24 ] ). Here each pass has its own threshold, which depends on the PCI at a significance level of 1% (Section S2.2).
Thresholds of the SHP tests
By § 2.2.4 , a fixed threshold rejects a fraction of the homogeneous candidates that increases with the temporal correlation of the speckle. The second-pass threshold of FaSHPS and the thresholds of CMP in this paper depend on the PCI η p \eta_p η p of the center (§ 2.2.2 ), which rises with the coherence (Table 1 ).
The table z α ( η ) z_\alpha(\eta) z α ( η ) of § 2.2.5 is built from simulated homogeneous clutter without fringes, with the stack size, search window and point spread function of the configuration, at the coherences c ∈ { 0 , 0.1 , 0.2 , 0.35 , 0.5 , 0.65 , 0.8 , 0.95 } c \in\{0, 0.1, 0.2, 0.35, 0.5, 0.65, 0.8, 0.95\} c ∈ { 0 , 0.1 , 0.2 , 0.35 , 0.5 , 0.65 , 0.8 , 0.95 } of Eq. (27) . For each coherence, η \eta η is the median of the PCI over the scene, and z α z_\alpha z α the 1 − α 1-\alpha 1 − α quantile of the second-pass statistic (Eq. (19) ) over the pixels outside the guard sets of 240 randomly placed centers. Both sequences are made monotone, interpolated linearly and held at their end values outside the tabulated range. The first-pass threshold z 1 z_1 z 1 is the median of the first-pass statistic at c = 0.5 c=0.5 c = 0.5 . The two thresholds of CMP are tabulated over η \eta η in the same way, from the same clutter and centers, for the decay constant of its coherence model.
The original thresholds of FaSHPS (Jiang et al. [6 ] , Eq. (7) and Section III-A) are normal quantiles of a 50 % interval in the first pass and of a 1 − α 1-\alpha 1 − α interval in the second, scaled with a coefficient of variation of the temporal mean amplitude estimated from the scene (§ 2.2.5 ). We estimate this coefficient of variation as Jiang et al. [6 ] , Sections III-B and IV, from standard deviations of the temporal mean amplitude, corrected by their Eq. (12) for the truncation of the selected set and regressed on the mean through the origin by iteratively reweighted least squares with Huber weights. The estimate here differs from theirs in two respects. The standard deviations are taken over the selected sets of the two passes in 2000 search windows at random centers, instead of in nonoverlapping blocks of 10 × 10 10 \times10 10 × 10 pixels. The interval is centered on the mean of the first pass, instead of on the amplitude of the block’s median pixel. In the simulated scenes, each of a single coherence, this estimate follows the coherence, from 1.44 times c A / N c_A/\sqrt{N} c A / N at c = 0 c=0 c = 0 to 5.84 times at c = 0.95 c=0.95 c = 0.95 (TSX-44, homogeneous scenes of Table S1 ). One value per scene then fits every center, so that Figure 3 shows the original thresholds in this favorable case. The original thresholds of CMP are the thresholds of its test in homogeneous clutter whose coherence follows the thermal and temporal terms of its coherence model without the scale 0.3, as Zhao et al. [24 ] (§ 2.2.1 ) simulate their data. Table S1 completes Figure 3 with the share of the candidates of the same class that each test rejects at a significance level of 1 %, for FaSHPS and CMP in the three configurations, in clutter of a single coherence and at coherence steps of equal amplitude and phase. § 2.2.5 gives the rejection by FaSHPS. At a coherence step, one threshold per scene cannot fit every center. There the original threshold of FaSHPS rejects as many of the candidates of the same class as in homogeneous clutter of the higher coherence or more, whereas the thresholds that depend on the PCI of the center reject about as many as in homogeneous clutter. CMP with its original thresholds rejects almost none of the homogeneous candidates up to c = 0.8 c=0.8 c = 0.8 , far fewer than its significance level, and 1.6 to 4.0 % at c = 0.95 c=0.95 c = 0.95 , compared with 1.1 to 1.2 % there with the thresholds that depend on η \eta η . Below c = 0.95 c=0.95 c = 0.95 the weak dependence of this rejection on the coherence agrees with Zhao et al. [24 ] , Section 3.1.1 and Figure 4 , who found their threshold only slightly affected by the coherence for two coherence matrices whose mean coherences differ by 0.14. The thresholds of CMP that depend on η \eta η , however, reject the fewest homogeneous candidates, 0.3 to 0.5 %, at c = 0.9 c=0.9 c = 0.9 , between the calibration coherences 0.8 and 0.95, where their interpolation is too permissive (three configurations).
Details of the truncation step
This section gives the rules of the truncation step (§ 2.2.6 ) that Algorithm A.1 uses beyond the main text. The ellipse order is replaced by the distance order where M p M_p M p is not positive definite, where the denominator of Eq. (20) is at most 1 % of the number of its terms, or where the smallest eigenvalue of M p M_p M p is at most 0.05 pixel2 ^2 2 . Ties, such as the two pixels at offsets d q \mathbf{d}_q d q and − d q -\mathbf{d}_q − d q , are kept in a fixed order. The candidate sizes L k = [ L 1 1.5 k − 1 ] L_k = [L_1\,1.5^{k-1}] L k = [ L 1 1. 5 k − 1 ] of Eq. (21) are used for every k k k with L 1 1.5 k ≤ ∣ O p ∖ G p ∣ L_1\,1.5^k \leq|O_p \setminus G_p| L 1 1. 5 k ≤ ∣ O p ∖ G p ∣ , followed by the last size L K = ∣ O p ∖ G p ∣ L_K = |O_p \setminus G_p| L K = ∣ O p ∖ G p ∣ . The truncation step is skipped if fewer than 38 members lie outside the guard set, i.e., if 1.5 L 1 > ∣ O p ∖ G p ∣ 1.5\,L_1 > |O_p \setminus G_p| 1.5 L 1 > ∣ O p ∖ G p ∣ , since then no size below L K L_K L K is a candidate.
Simulation study
This section completes § 2.3 . In this section, a value or a range refers to all centers of the evaluation scenes, the full window, FaSHPS and CMP, and the three configurations TSX-44, S1-79 and S1-20, unless the sentence or the parentheses after it name others. A range gives the smallest and the largest value over them.
Tuning and checks of the design
The criterion of the tuning of § 2.3.1 is the weighted geometric mean of the phase linking error relative to the oracle over all centers of the tuning scenes, averaged over the SHP selections with ANS and over the configurations. The tuning has three steps, each of which varies one parameter with the others held fixed, and a change counts only if it lowers the criterion by more than 0.1 %. First, at κ = 0.75 \kappa= 0.75 κ = 0.75 and λ = 3 \lambda= 3 λ = 3 , ρ min \rho_{\min} ρ m i n takes the value of the lowest criterion, or the largest value within 0.1 % of it. Second, at this ρ min \rho_{\min} ρ m i n , κ \kappa κ moves from 0.75 to a neighboring tested value, and further in the same direction, as long as each move counts. Third, at this κ \kappa κ , λ \lambda λ moves from 3 by the same rule. In addition, λ \lambda λ was varied at κ = 0.75 \kappa= 0.75 κ = 0.75 for ρ min = 0.02 \rho_{\min} = 0.02 ρ m i n = 0.02 , 0.03, 0.04 and 0.05, and every tested value from 0 to 4 changed the criterion by less than 0.1 % against λ = 3 \lambda= 3 λ = 3 . The tuning selected ρ min = 0.025 \rho_{\min} = 0.025 ρ m i n = 0.025 , κ = 0.5 \kappa= 0.5 κ = 0.5 and λ = 3 \lambda= 3 λ = 3 . At κ = 0.75 \kappa= 0.75 κ = 0.75 and λ = 3 \lambda= 3 λ = 3 , the criterion at ρ min = 0.025 \rho_{\min} = 0.025 ρ m i n = 0.025 is 2.1 % lower than at ρ min = 0.05 \rho_{\min} = 0.05 ρ m i n = 0.05 , and that at ρ min = 0.02 \rho_{\min} = 0.02 ρ m i n = 0.02 lies within 0.04 % of it. At ρ min = 0.025 \rho_{\min} = 0.025 ρ m i n = 0.025 , the criterion differs by at most 1.0 % over the tested values of κ \kappa κ .
Table S2 gives the criterion on the tuning scenes, averaged over the methods and configurations, for each bound ρ min \rho_{\min} ρ m i n and each tested setting of the margin factor κ \kappa κ and the noise multiplier λ \lambda λ , and Table S3 the criterion of each method at the selected ρ min \rho_{\min} ρ m i n and λ \lambda λ . One value of each factor applies to every method and configuration. At the tuned ρ min = 0.025 \rho_{\min} = 0.025 ρ m i n = 0.025 , g ( ρ min ) 2 ≈ 0.00039 g(\rho_{\min})^2 \approx0.00039 g ( ρ m i n ) 2 ≈ 0.00039 , which is 0.8 to 3.0 % of the standard deviation of one corrected SPC value under independent uniform phases at N = 20 N = 20 N = 20 to 79 (Eq. (36) ), so that the speckle that two pixels beyond the speckle correlation length still share adds for less than the noise of one corrected SPC value. For ρ min \rho_{\min} ρ m i n between 0.024 and 0.025, the core, the guard set, the ring R p \mathcal{R}_p R p and the pixels of the ellipse order are unchanged on both sensors, so that the tuning fixes ρ min \rho_{\min} ρ m i n only to within this interval.
A smaller margin does not remove the ensembles that remain too large (limitation (iii) of § 3.2 ).
On the tuning scenes at the tuned bound ρ min = 0.025 \rho_{\min} = 0.025 ρ m i n = 0.025 , the smallest margin factor tested there, κ = 0.375 \kappa= 0.375 κ = 0.375 , was compared with the tuned κ = 0.5 \kappa= 0.5 κ = 0.5 (ANS, full window, FaSHPS and CMP, TSX-44 and S1-79). The smaller factor lowered the share of the area at distributed centers whose ensemble has more than four times the oracle size by 0.6 to 1.1 percentage points, to 11 to 16 %. The share of truncated ensembles rose to 39 to 54 %, compared with 36 to 50 % at κ = 0.5 \kappa= 0.5 κ = 0.5 . A smaller margin thus truncates more ensembles but hardly reduces the share of those that remain too large, so that the limitation lies in the noise of the score. With ANS, the phase linking error of FaSHPS exceeds twice the oracle error at 13 % (TSX-44) and 9.7 % (S1-79) of the area of the evaluation scenes, with the truncation step alone at 14 % (TSX-44) and 8.9 % (S1-79), and for FaSHPS alone at 40 % (TSX-44) and 27 % (S1-79).
The fixed values, those of Table 2 and the other constants of § 2.2 , were set without tuning. Table S4 and Table S22 give the change of the error when the significance level, the percentile of the null quantile η 0 \eta_0 η 0 , n min n_{\min} n m i n , L min L_{\min} L m i n , R R R , the factor of the candidate sizes, the tolerance, the ring or the 5 × 5 5 \times5 5 × 5 window of the PCI is changed. The factor 1.5 of the candidate sizes trades the resolution of the size against the noise of the maxima. The ring of 24 pixels has the size of a 5 × 5 5 \times5 5 × 5 neighborhood without its center.
Table S22. Sweep of the fixed constants of ANS (Table 2 and § 2.2 ) and of the tuned bound ρ min \rho_{\min} ρ m i n , one at a time. Table S18 compares further sets of reference acquisitions for FaSHPS, and Table S2 gives the tuning of ρ min \rho_{\min} ρ m i n on the tuning scenes. Phase linking error over all centers of the evaluation scenes relative to that of the same method with ANS at the values of the main comparison, ratio of the weighted geometric means on the same centers. Each row replaces one of the values of the main comparison, namely n min = 100 n_{\min}=100 n m i n = 100 , R = 8 R=8 R = 8 , the ratio of 1.5 between consecutive candidate sizes, the tolerance ζ / 3 \zeta/3 ζ /3 , L min = 25 L_{\min}=25 L m i n = 25 , the ring of 24 pixels, the average of η ~ \tilde{\eta} η ~ over 5 × 5 5\times5 5 × 5 pixels and ρ min = 0.025 \rho_{\min}=0.025 ρ m i n = 0.025 . The ring, the average of η ~ \tilde{\eta} η ~ and ρ min \rho_{\min} ρ m i n change the PCI, and with them η 0 \eta_0 η 0 and the thresholds of FaSHPS are calibrated again.
Table S2. Tuning of the truncation step on the tuning scenes: criterion (Section S3.1, column All of Table S3 ) for each speckle correlation bound ρ min \rho_{\min} ρ m i n and each setting of the margin factor κ \kappa κ and the noise multiplier λ \lambda λ that was run. The tuning chose ρ min \rho_{\min} ρ m i n at κ = 0.75 \kappa= 0.75 κ = 0.75 and λ = 3 \lambda= 3 λ = 3 (first column), then κ \kappa κ at the chosen ρ min \rho_{\min} ρ m i n with λ = 3 \lambda= 3 λ = 3 , then λ \lambda λ at the chosen κ \kappa κ , and accepted a change only where it lowered the criterion by more than 0.1 %. The three columns right of the vertical rule show the effect of λ \lambda λ at κ = 0.75 \kappa= 0.75 κ = 0.75 for the other bounds, and κ = 0.25 \kappa= 0.25 κ = 0.25 was run at ρ min = 0.02 \rho_{\min} = 0.02 ρ m i n = 0.02 only. A dash marks a setting that was not run, and * the selected setting.
Table S3. Tuning of the truncation step on the tuning scenes at the tuned bound ρ min = 0.025 \rho_{\min}=0.025 ρ m i n = 0.025 . Weighted geometric mean of the phase linking error relative to the oracle over all centers of each method with the PCI-scaled window and the gated truncation step, for the margin factor κ \kappa κ at the selected noise multiplier λ = 3 \lambda=3 λ = 3 , and the criterion, its geometric mean over the methods and configurations (All). Full is the full window, and the first row (none) has neither the truncation step nor the PCI-scaled window. ∗ ^{*} ∗ marks the selected κ \kappa κ .
The SHP selections use the settings of § 2.2 , except the decay constant of the coherence model of CMP, which was fitted to the mean coherence of the single-reference interferograms over the first ten scenes of each batch of tuning scenes. The mean coherence of an interferogram is the square root of the mean of Re ( v p 1 n v q 1 n ‾ ) \operatorname{Re}(v_p^{1n}\overline{v_q^{1n}}) Re ( v p 1 n v q 1 n ) over all pixels p p p and the pixels q q q of their rings R p \mathcal{R}_p R p , mapped through the inverse of g g g . The fit gave 2271, 6713 and 1044 days (TSX-44, S1-79 and S1-20, respectively).
Table S4 gives checks of the design on the evaluation scenes, each as the ratio of the error over all centers to that of the same method with ANS. The truncation step gated by η \eta η on the search window, without the PCI-scaled window, has 1.00 to 1.02 times the error of ANS, so that the PCI-scaled window hardly changes the error over all centers and pays off in time (§ 3.1.4 ). Conversely, a window that shrinks much faster than the PCI-scaled window, without the truncation step, tests whether a smaller window alone could replace the truncation step. Its semiaxes shrink linearly with the PCI, from the search window at η 0 \eta_0 η 0 to the smallest ellipse of 25 pixels, the minimum ensemble size, already at the median PCI of homogeneous clutter with c = 0.95 c = 0.95 c = 0.95 . Those of the PCI-scaled window reach the 100 nearest pixels only at η p = 1 \eta_{\mathrm{p}} = 1 η p = 1 (§ 2.2.3 ). This window has 1.132 to 1.192 times the error of ANS (full window and FaSHPS), so that a smaller window alone cannot replace the truncation step. The PCI-scaled window alone, without the truncation step, is the variant +S of Table 3 . The distance order and the margin 1.5 / N 1.5/N 1.5/ N have a higher error than the ellipse order and the jackknife margin. For the distance order the excess lies mostly at the centers with four to eight fringes per window diameter (1.01 to 1.05 times). The percentile of the null quantile η 0 \eta_0 η 0 changes the error by at most 0.3 %. The rows on the significance levels and the original thresholds of the SHP tests are discussed in § 2.3.2 .
Ordinary sample covariance
This section completes § 2.2.8 . With the ordinary sample covariance in place of Eq. (26) , and both relative to the oracle of Eq. (26) (Table S5 ), the error over the distributed centers is higher for every SHP selection, with the truncation step and with ANS. It is at most 7.6 % higher for FaSHPS and CMP and about twice as high for the full window (1.78 to 1.88 times). A few bright pixels dominate the sample covariance of a large ensemble, and the SHP tests exclude them, which is why FaSHPS and CMP lose least. We use Eq. (26) throughout.
Table S5. Amplitude-normalized estimator (Eq. (26) ) compared with the ordinary sample covariance: phase linking error relative to the oracle of the amplitude-normalized estimator, weighted geometric mean over the distributed centers of the evaluation scenes. Columns with the truncation step (+T) and with ANS as in Table 3.
Scene generator
Table S6 lists the components of the random scenes of § 2.3.1 .
Phase linking solver
The objective J J J of Eq. (28) in § 2.3.1 is maximized from the phases of the leading eigenvector of ∣ Γ ^ ∣ ⊙ Γ ^ |\hat{\Gamma}| \odot\hat{\Gamma} ∣ Γ ^ ∣ ⊙ Γ ^ . Up to ten sweeps of cyclic coordinate ascent come first, each of which sets one phase after the other to its conditional maximum. Damped Newton iterations follow until the largest component of the gradient falls below 10 − 6 10^{-6} 1 0 − 6 times the sum of the weights. If the Newton iterations do not converge within 100 iterations, coordinate ascent continues from their last iterate for at most 1000 sweeps, after which the last iterate is kept. On the evaluation scenes, every solution converged. The same solver is used for the real data.
Absolute errors
Table S7 gives the absolute errors of FaSHPS in each coherence class.
Table S7. Weighted median oracle error and phase linking error in degrees of FaSHPS alone, with the truncation step (T) and with the PCI-scaled window and the gated truncation step (ANS), evaluation scenes. Rows as in Table 3.
Other summaries of the error
The weighted arithmetic mean and the weighted root mean square of the phase linking error over all centers give more weight to the centers with large errors than the weighted geometric mean of the ratio to the oracle (§ 2.3.1 ). With either summary, ANS lowers the error relative to the SHP selection alone, and the largest window with ANS has the lowest error of the four windows and less error than the same SHP selection without either step at its best window, as with the geometric mean. The reductions are smaller with the root mean square. With the arithmetic mean and the root mean square, respectively, ANS lowers the error at the search window of Table 2 by 18 to 26 % and by 10 to 15 %, against 21 to 31 % with the geometric mean. The largest window with ANS has 7.9 to 11 % and 5.4 to 8.4 %, respectively, less error than the same SHP selection without either step at its best window, against 9.3 to 11 % (three configurations).
The 95 % intervals of a bootstrap resampling of the scenes [2 ] , paired between the methods, have half-widths of at most 0.015 for the ratios between the SHP selections over all centers and for those between the variants of one method. The intervals include one only for ANS against the truncation step alone in S1-79, and for CMP against the full window in TSX-44 with the truncation step alone and with ANS. The other ratios thus differ from one by more than the sampling noise of the scenes.
Gate
This section completes the paragraph on the contributions of the two parts in § 2.3.2 . At c = 0 c = 0 c = 0 , the gate passes 0.8 to 1.5 % of the distributed centers farther than 20 pixels from a point scatterer and 0.5 window radii from a class border. It passes 66 to 96 % of those closer than three pixels to a scatterer and 47 to 73 % of those within 0.05 window radii of a border. The gate thus rarely applies the truncation step in zero-coherence clutter far from scatterers and borders, and often near them. The region farther than 20 pixels from a point scatterer and 0.5 window radii from a class border holds 31 to 33 % of the area at c = 0 c = 0 c = 0 , so that most zero-coherence area lies near scatterers or borders. At a coherence of 0.2, the error relative to the oracle with ANS differs from that with the truncation step alone by at most 0.040, so that at low coherence the gate changes the error little. At c = 0.95 c = 0.95 c = 0.95 the median of η \eta η falls with the fringe density but stays far above η 0 \eta_0 η 0 , at 0.68 to 0.77 at four or more fringes per window diameter, so that the gate passes almost every coherent center in dense fringes (at least 99.9 % of the distributed centers with c ≥ 0.5 c \ge0.5 c ≥ 0.5 in every fringe-density bin). At two or more fringes per window diameter, fewer FaSHPS ensembles are truncated with ANS than with the truncation step alone (75 to 76 % against 79 to 80 %, three configurations).
Size of the search window
This section completes the paragraph on the size of the search window in § 2.3.2 . Table S8 gives the error over all centers for the four windows of Figure 6 .
In zero-coherence clutter (c = 0 c = 0 c = 0 ), the largest window has the lowest error of the four, for the SHP selections alone, with the truncation step and with ANS. There the smallest window has 1.24 to 1.40 times the error of the largest, without and with the truncation step.
Over all centers and with ANS, the advantage of FaSHPS and CMP over the full window is smallest at the smallest window. There CMP+ANS even has 2.2 % more error than the full window with ANS in TSX-44 (Table S8 ).
The truncation step and ANS also lower the error when every variant is used at its own best window, chosen in hindsight on the evaluation scenes. Each SHP selection then has 9.3 to 11 % less error with the better of the two than without either step.
Even at the window where the SHP selection alone has its lowest error, chosen in hindsight on the evaluation scenes, the truncation step lowers the error over all centers by 4.5 to 7.5 %. For FaSHPS it lowers the error at the distributed centers with one fringe or more per window diameter by 15 to 24 %. ANS lowers the error at these centers less than the truncation step alone, since the PCI-scaled window shrinks in proportion to the search window.
In coherent clutter with few fringes, a truncation raises the error at the smallest window and lowers it at the largest. At the distributed centers of coherence 0.5 to 0.8 with fewer than 0.25 fringes per diameter of the largest window, the error of the truncated FaSHPS ensembles rises by 10 to 18 % at the smallest window and falls by 22 to 28 % at the largest. At the largest window more of the ensembles are truncated, and to a smaller share of their size. The smallest windows hold 749 pixels (TSX-44) and 535 pixels (S1-79 and S1-20), and there the loss of looks by a truncation raises the error more than the removal of bias lowers it.
With ANS, the error of the four windows differs by only 1.9 to 7.7 % both at the centers with one fringe or more per window diameter and at those near discontinuities, where the truncation step acts. At c = 0 c = 0 c = 0 and c = 0.2 c = 0.2 c = 0.2 the smallest window has the largest error.
The largest window with ANS keeps its advantage when the centers are weighted so that the distribution of η \eta η matches that of a real stack, a weighting that ignores the fringe density. It then has 6.2 to 9.5 % less error than the same SHP selection without either step at its best window. This weighting uses the real stacks of Section 5, namely Visp, the slow-deformation, Maiun and long-term coherent test areas and the landslide substack, each with its configuration. It is applied to the full window, FaSHPS and CMP.
Elementary scenes
This section completes the paragraphs on where ANS lowers the error and on the contributions of its two parts in Section 4.2. Table S9 gives the error of every variant in the elementary scenes of TSX-44. In homogeneous clutter, every SHP selection alone is within 2.9 % of the oracle error at all three coherences (TSX-44), and the truncation step and ANS raise the error, since ensembles are truncated there only by chance. In coherent clutter the multilooked phase of a small ensemble is already precise, so that the scores of the candidate sizes differ little, the maximum falls on a smaller size by chance, and its excess over the score of the whole ensemble exceeds the margin more often. At c = 0.5 c = 0.5 c = 0.5 , 17 to 39 % of the ensembles are truncated with the truncation step, although homogeneous clutter needs no truncation. In absolute terms, for the full window in TSX-44, the weighted median phase linking error is 4.11°, 1.77° and 0.42° at c = 0.2 c = 0.2 c = 0.2 , 0.5 and 0.95, respectively, without the truncation step and 4.20°, 1.99° and 0.61° with it, compared with 4.02°, 1.76° and 0.42° at the oracle size. The added error lies in the truncated ensembles, and the 90th percentile rises from 5.1°, 2.2° and 0.5° to 7.9°, 5.3° and 1.8°. The error that the truncation step adds in homogeneous clutter is thus small in degrees.
At a phase step of 3 rad and c = 0.95 c = 0.95 c = 0.95 , every SHP selection has 1.87 to 3.09 times the oracle error, alone, with the truncation step and with ANS (TSX-44). Tests on amplitudes do not see a boundary in phase, and CMP gives the phase differences little weight, since its assumed coherence is at most 0.3 (Eq. (37) ). An ensemble is truncated at such a boundary only where the boundary is close to the center. With an additional amplitude contrast of 6 dB at a phase step of 3 rad and c = 0.65 c = 0.65 c = 0.65 , the SHP tests detect the boundary in part, CMP better than FaSHPS, and the truncation step raises the error of CMP there. At the coherence step, the SHP selections alone mix the two sides with no loss in phase, the truncation step raises the error about as in homogeneous clutter at c = 0.5 c = 0.5 c = 0.5 , and ANS raises it further, since it also reduces the window at the coherent centers.
Further results on the truncation step
For the three SHP selections at c = 0.95 c = 0.95 c = 0.95 , ANS raises the error relative to the truncation step alone in the scenes without atmospheric screen, by 11 to 13 % in TSX-44 and by 28 to 30 % in S1-79. In the scenes with a screen it lowers the error, by 15 % in TSX-44 and by 10.1 to 10.2 % in S1-79.
With few fringes, the truncation step lowers the error most where the window contains a class border, a discontinuity or a strong atmosphere. Below 0.25 fringes per window diameter and at c ≥ 0.8 c \ge0.8 c ≥ 0.8 , it lowers the error of the full window by 34 to 43 % where a class border or a discontinuity lies within one window radius of the center or the atmosphere is strong. These centers hold 68 to 73 % of the area below 0.25 fringes per window diameter and at c ≥ 0.8 c \ge0.8 c ≥ 0.8 . At the other centers the truncation step lowers the error by 7.1 to 11 %. Over all centers within 0.25 window radii of a discontinuity, at any fringe density, the error of the full window falls from 2.40 to 1.43 times the oracle error, and over all centers of the scenes with the strong atmosphere from 2.28 to 1.33 (TSX-44).
Below 0.25 fringes per window diameter and at c ≥ 0.5 c \ge0.5 c ≥ 0.5 , the weighted median phase linking error of the full window is 6.2° to 8.4°, and the truncation step lowers it to 5.0° to 6.4°. At few fringes, the error and its reduction by the truncation step are thus small in degrees.
Further results on the SHP selections
Without either step, FaSHPS accepts 81 to 89 % and CMP 73 to 81 % of the search window on average over all centers, and within the PCI-scaled window of ANS they accept about the same shares of it. With and without ANS, at most 0.5 % (FaSHPS) and 1.9 % (CMP) of all centers are insufficient, compared with none for the full window.
The other two errors follow the phase linking error. The coherence error (Eq. (29) ) over the distributed centers is 0.124 to 0.146 for FaSHPS alone, 0.080 to 0.106 with the truncation step and 0.075 to 0.101 with ANS. Table S10 gives the matrix error of the main comparison. Relative to the matrix error at the oracle size, it is 1.34 to 1.56 over all centers for the SHP selections alone and 1.06 to 1.14 with the truncation step or ANS.
Experiments on TerraSAR-X and Sentinel-1 data
This section completes § 3 .
Processing parameters
Table S11 gives the parameters of Eq. (11) for the speckle correlation in § 2.1.5 and § 3.1.1 .
The TerraSAR-X stack over Visp has a single polarization (horizontal transmit and receive), a descending orbit and right-looking geometry, with an incidence angle of about 45 ∘ 45^\circ 4 5 ∘ (43.3 ∘ 43.3^\circ 43. 3 ∘ to 45.8 ∘ 45.8^\circ 45. 8 ∘ across the swath). The X-band data has a range bandwidth of 100 MHz and a pixel spacing of 1.36 m in slant range and 1.98 m in azimuth. Phases are referenced to the acquisition of 11 September 2019, and perpendicular baselines relative to it range from − 206 -206 − 206 to 128 m.
The decay constant of the coherence model of CMP, fitted to each real stack, is 3349 days for Visp, and 3118, 1778, 3292 and 602 days for the slow-deformation, Maiun and long-term coherent test areas and the landslide substack, respectively. These constants, like those fitted in the simulations (Section S3.1), are much longer than the 200 days that Zhao et al. [24 ] set. The mean coherence of the single-reference interferograms falls over the first part of each stack and then levels off, as the coherence of Eq. (27) does in the simulations, and an exponential with a free level, fitted over the whole stack, follows the slow part, which gives the long constants.
For Sentinel-1, Eq. (11) predicts a speckle correlation of 0.82, 0.42 and 0.11 at range lags one to three, and 0.66, 0.17 and 0.04 in azimuth. Compared with ideal band-limited oversampling, the B-spline interpolator that oversamples the data in range (Section S4.4) raises the range values by at most 0.005. The reference acquisition of the coregistration matches this prediction to within 0.007 at range lags one to three in the Maiun and long-term coherent test areas, measured per acquisition as the magnitude of the summed products of neighboring samples relative to the summed intensity. In the 78 coregistered acquisitions, the correlation at these lags exceeds the prediction by up to 0.042, and the sidelobes are damped, with a median of 0.026 at range lag five in both test areas against 0.043 predicted. Both deviations are of the size that the resampling of the coregistration produces. The coregistration resamples with a Lanczos interpolator of order 4, and at fractional range shifts of up to half a pixel, this interpolator followed by the B-spline raises the correlation at lags one to three by up to 0.040 and gives 0.020 to 0.039 at lag five.
Mask of the Visp stack
This section completes the comparison of the NSPC mask with the layover and shadow maps of § 3.1.2 (Figure 8 ). The shares in this section refer to the valid pixels, since pixels of the Visp scene with a zero value in at least one acquisition, 0.9 % of the scene, are excluded from the outset (§ 2.2.1 ). Table S12 gives the shares at h = 0.25 h = 0.25 h = 0.25 , where the mask covers 56 % of the layover of the terrain model.
Table S12. NSPC mask of Visp (h = 0.25 h = 0.25 h = 0.25 , after topographic phase removal) against the layover and shadow maps of Figure 8 , computed from the terrain model swissALTI3D and from the surface model swissSURFACE3D Raster. A pixel is masked if its NSPC lies below h h h , and its class is the one that the map assigns to it. The first block gives how the masked pixels divide among the classes of a map, the second the share of all pixels that a map puts in layover or shadow, and the third the share of the pixels of each class that are masked. Shares refer to the valid pixels, and the town section is the dashed box of Figure 8 . A dash marks a value that was not computed.
In the layover of the terrain model, the median NSPC is 0.45 before topographic phase removal, at the correlated-noise floor, and 0.21 after it, and 0.2 % of these pixels lie below h = 0.25 h = 0.25 h = 0.25 before the removal. In its shadow the median NSPC is 0.39 before and after the removal, close to the correlated-noise floor of 0.43.
Over the whole scene, 54 % of the masked pixels lie outside both classes of the terrain model map (Table S12 ). Their low agreement is a property of the data and does not arise from the topographic correction, since 93 % of them are masked before it as well. They are dark, with a median amplitude of − 6.6 -6.6 − 6.6 dB relative to the scene median, similar to the shadow of the terrain model (− 7.1 -7.1 − 7.1 dB) and unlike its layover (+9.5 dB), which argues against building layover as the main cause. They also lie close to bright pixels, i.e., the 2 % of pixels with the highest mean amplitude, at a median distance of 6 pixels, compared with 21 for all pixels. Dark pixels, i.e., those more than 3 dB below the scene median, are masked more often the closer they lie to a bright pixel, in the town section (4.8 % beyond 16 pixels and 22 % within 4 pixels) as in the rest of the scene (2.0 % and 18 %, respectively).
The surface model map places most of the masked pixels outside both classes of the terrain model map in its shadow. Its classes agree with the mean amplitude, with a median of +2.6 dB in layover and − 4.0 -4.0 − 4.0 dB in shadow relative to the scene median. Its layover contains 94 % of the layover of the terrain model. Of the masked pixels outside both classes of the terrain model map, it places 57 % in shadow and 8.5 % in layover.
Outside the layover of the terrain model, the masking of the surface model classes hardly depends on the removal of the topographic phase. Over the whole scene, 0.4 % of the surface model layover outside that of the terrain model lies below h h h before the removal and 0.5 % after, while the masked share of the surface model shadow in the town section does not depend on it (11 % before, 11 % after).
Within the town section the mask covers the layover of the surface model least (Table S12 ), since bright returns keep NSPC high there (median 0.49), and being in shadow is not sufficient for a pixel to be masked. Of the shadow pixels in the town section more than 16 pixels from a bright pixel, 4.2 % are masked, compared with 14 % at 2 to 8 pixels. Of the masked pixels of the town section in neither class of the surface model map, 48 % lie within two pixels of shadow. Most masked town pixels thus lie in or beside building shadow near bright pixels. Over the forest, the surface model classes follow the canopy at the scale of single trees, and at most 1.7 % of the pixels of any class are masked.
Point scatterers in clutter with zero coherence
Table S13 gives the simulation of Section 5.3, in which NSPC is measured at the clutter pixels. Within two to six pixels of dense scatterers with different heights, 25 to 40 dB above the clutter (third row of Table S13 ), the median NSPC decreases to 0.33 to 0.37, and 14 to 16 % of the clutter pixels fall below h = 0.25 h = 0.25 h = 0.25 . Among the dark pixels of the town section of Visp within four pixels of a bright pixel, 22 % are masked, against 14 to 16 % of the simulated clutter pixels near dense bright scatterers. The decrease follows from the sign of the sidelobes. Beyond the main lobe, the point spread function alternates in sign from one sample to the next along range, at up to − 33 -33 − 33 dB relative to the peak. For a sidelobe amplitude equal to that of the clutter, the expected value of Q + \sqrt{Q^{+}} Q + for two range neighbors falls from 0.49 for clutter alone to 0.30 with a sidelobe of opposite sign, and with equal signs it would rise to 0.75. With scatterers of 10 to 25 dB the decrease of NSPC is confined to a distance of two pixels, so a wider decrease requires scatterers 25 dB or more above the clutter. By the sign of the sidelobes (Section 5.3), a single scatterer is sufficient to lower NSPC below the floor. Several scatterers with different phase histories, as left by buildings of different height that the terrain model does not contain, increase the effect. With scatterers of 25 to 40 dB, the masked share is 2.0 to 2.8 times that of scatterers with one common phase history. Around a scatterer of 10 to 25 dB or an isolated scatterer, NSPC is lowest close to the scatterer and rises toward the floor with the distance (first, second and last rows of Table S13 ). On the dense grid of scatterers of 25 to 40 dB, a pixel 5 to 6 pixels from the nearest scatterer lies near the middle between four of them, where their sidelobes overlap, and NSPC falls with the distance (third and fourth rows) and is lower with different heights than with one.
Table S13. Simulated NSPC of clutter pixels with zero coherence near point scatterers (TSX-44, N = 44 N = 44 N = 44 ). Entries are the median and the share below h = 0.25 h = 0.25 h = 0.25 , by distance to the nearest scatterer. The correlated-noise floor is 0.43. Scatterers lie on a grid with a spacing of 8 pixels (dense) or 40 pixels (isolated), each shifted by up to two pixels along each axis at random. Their phase histories follow from heights between 0 and 20 m, or from one common height.
Sentinel-1 site
This section completes the description of the Sentinel-1 site of § 3.1 .
The scene of § 3.1 runs from Rahimabad (36.12° N, 74.30° E) to Shishkat (36.36° N, 74.87° E) on Attabad Lake. The relief is extreme. Within the scene the valley floor rises from about 1500 m to about 2200 m, and the terrain reaches about 7500 m on the north flank of Rakaposhi (7788 m). The slopes are arid and mostly bare rock and debris, with vegetation confined to irrigated terraces around the villages, so that large parts of the scene remain coherent over years, while the steep slopes cause extended layover and shadow. Slope instability is widespread. Landslide inventories along the highway, compiled or updated with InSAR, list hundreds of landslides ([5 , 18 ] ), among them slow-moving landslides of glacio-fluvial and moraine deposits above the villages of the Hunza–Nagar section, the fastest with a line-of-sight rate of more than 0.1 m per year ([20 ] ).
The acquisitions are from subswath IW1, with vertical transmit and receive polarization, on ascending orbits. The stack was coregistered to 11 May 2024 with the coregistration procedure of the GAMMA software for data of the terrain observation with progressive scans (TOPS) mode ([22 ] ), with the temporally adjacent acquisition as secondary reference for the spectral diversity, and deramped. The oversampling in range uses a B-spline interpolator of degree 9, and the pixel spacing of the oversampled data is 1.16 m in slant range and 13.97 m in azimuth.
Sentinel-1 test areas
Table S14 lists the test areas of § 3.1 .
Sentinel-1 mask
Table S15 gives the comparison of the NSPC mask with the geometric layover and shadow map of § 3.1 . In the four test areas used for the mask, shadow lies 7.3 to 9.8 dB below the median intensity of its test area, and layover 8.5 to 12.6 dB above it, as expected from the imaging geometry.
On the native range grid, with every second range sample, the predicted lag-one range correlation is 0.42, against 0.82 on the oversampled grid. Simulated clutter with zero coherence then has a median NSPC of 0.32, against 0.54 on the oversampled grid (S1-79), so that 27 % of it falls below h = 0.30 h = 0.30 h = 0.30 . At h = 0.20 h = 0.20 h = 0.20 , the 0.1 % quantile of NSPC in this clutter, the mask covers 35 to 63 % of layover and 54 to 83 % of shadow in the four test areas used for the mask. It also covers 13 to 25 % of the pixels in neither class, against 6.3 to 13.8 % at h = 0.30 h = 0.30 h = 0.30 on the oversampled grid.
Before topographic phase removal, the mask covers 0.6 to 4.4 % of layover instead of 31 to 63 %. In shadow the median NSPC is then 0.29 to 0.35, and the masked share of shadow changes by at most 2.6 percentage points. Pixels in neither class with the same temporal mean intensity as shadow have, before the removal, a similarly low NSPC, and those with the intensity of layover a higher NSPC than layover. The low NSPC of shadow thus goes with its low intensity, and layover agrees with its neighbors slightly less than other pixels of the same intensity (median NSPC 0.55 to 0.56 against 0.59 to 0.67 before the removal). In the layover and shadow test area, 58 % of all masked pixels belong to one of the two geometric classes, and many of the remaining ones border them.
Further results on the phase-linked maps
On Visp the gate passes 39 % of the centers, and with ANS 16 % of the FaSHPS ensembles are truncated, 50 % of them in the town section, which holds 19 % of the centers. On the slow-deformation test area the gate passes 96 % of the centers and 40 % of the FaSHPS ensembles are truncated, and over the section of the landslide shown in Figure 12 the gate passes 90 % of the centers and 50 % of the FaSHPS ensembles are truncated. Of the centers of the Maiun test area, 5.8 % lie below the null quantile η 0 = 0.050 \eta_0 = 0.050 η 0 = 0.050 of S1-79.
On the slow-deformation test area, the median ensemble size of the full window falls from 4238 to 1715 pixels with ANS, and over the landslide section 30 % of the truncated FaSHPS+ANS ensembles keep less than a tenth of the FaSHPS ensemble at the same center. On the long-term coherent test area, where coherent clutter is uniform over large areas, ANS shrinks the search window at 97 % of the unmasked centers, to a median of 3005 of the 4437 pixels of the search window.
Below the null quantile, where FaSHPS+ANS returns the ensembles of FaSHPS, the truncation step alone would truncate 5.1 % of the FaSHPS ensembles on Visp (η 0 = 0.066 \eta_0 = 0.066 η 0 = 0.066 ) and 11 % on the Maiun test area, compared with 7.8 % and 5.0 %, respectively, at the simulated distributed centers with η < η 0 \eta< \eta_0 η < η 0 (TSX-44 and S1-79).
The fringe density on the slow-deformation slope is measured as in § 2.3.1 , as the phase gradient times the window diameter over 2 π 2\pi 2 π , in the normalized coordinates of the elliptic search window of 25 × \times × 225 pixels. The gradients are the angles of the means of the complex products of neighbors over 9 × \times × 81 pixels in azimuth and in range. The estimate of FaSHPS+ANS for the interval of Figure 11 has at least one fringe per window diameter at 21,081 centers with η ≥ 0.2 \eta\ge0.2 η ≥ 0.2 . At 78 % of them, the single look of the same interval averaged over 3 × \times × 27 pixels, which does not depend on ANS, has at least one as well, against 18 % of all centers of the test area with η ≥ 0.2 \eta\ge0.2 η ≥ 0.2 .
Acceptance of the SHP tests on the real stacks
FaSHPS and CMP accept fewer candidates on the real stacks than in the simulated scenes (§ 3.1.3 ). Table S16 gives the acceptance, i.e., the share of the unmasked pixels of the search window in the ensemble (of all its pixels in the simulations, which have no mask), by the PCI of the center. It also gives the second-pass threshold z α ( η p ) z_{\alpha}(\eta_p) z α ( η p ) of FaSHPS, the spread of the temporal mean amplitudes over the search window and the contrast in mean intensity that CMP tests. The spread is the robust standard deviation (1.4826 times the median absolute deviation) of A ˉ q \bar{A}_q A ˉ q over the unmasked pixels of the search window outside the guard set, in units of c A A ~ / N c_A \tilde{A}/\sqrt{N} c A A ~ / N , the scale of T p q T_{pq} T pq in Eq. (19) with the median A ~ \tilde{A} A ~ of A ˉ q \bar{A}_q A ˉ q in place of μ p \mu_p μ p . Where the spread exceeds z α z_{\alpha} z α and μ p \mu_p μ p is close to A ~ \tilde{A} A ~ , candidates at one standard deviation from A ~ \tilde{A} A ~ fall outside the interval of the second pass. The contrast is the median over the same pixels of ∣ 10 log 10 ( I ˉ q / I ˉ r ) ∣ |10\log_{10}(\bar{I}_q/\bar{I}_r)| ∣10 log 10 ( I ˉ q / I ˉ r ) ∣ , where I ˉ r \bar{I}_r I ˉ r , the reference of the second pass of CMP, is the mean intensity of its reference set.
Table S16. Acceptance of FaSHPS and CMP by the PCI η p \eta_p η p of the center on the real stacks and in the simulated scenes of their configuration, with the share of the centers in each bin, the second-pass threshold z α z_\alpha z α of FaSHPS at the edges of the bin and the spread and the contrast. The line of each configuration gives η 0 \eta_0 η 0 and the PCI from which z α ≥ N / c A z_\alpha\ge\sqrt{N}/c_A z α ≥ N / c A . For each center, the spread is the robust standard deviation of the temporal mean amplitudes over the unmasked pixels of its search window outside the guard set, relative to the scale of the FaSHPS statistic, and the contrast the median absolute difference in mean intensity, in dB, between the pixels of its search window and the reference set of CMP (Section S4.8). The table gives the median of each over 120 centers per bin of a real scene and over the centers of 30 evaluation scenes of a configuration, for the spread over the distributed centers of the texture class of the row. Shares and acceptance are taken over all centers, and every value of the simulated scenes is weighted as in § 2.3.1 .
Most Visp centers have a PCI below 0.3 (93 %, against 59 % in the simulated scenes of TSX-44), and there the mean amplitudes of Visp spread as much as in simulated clutter with strong texture or more. FaSHPS accepts 42 % of the unmasked pixels of the search window on Visp. Weighted with the shares of the PCI bins of Visp, the acceptance in the simulated scenes is 74 % instead of 85 %, so that most of the difference arises within the bins. Above a PCI of 0.24 to 0.31, depending on the configuration, z α z_{\alpha} z α exceeds N / c A \sqrt{N}/c_A N / c A . The interval μ p ( 1 ± z α c A / N ) \mu_p(1 \pm z_{\alpha}c_A/\sqrt{N}) μ p ( 1 ± z α c A / N ) of the second pass then includes every mean amplitude between zero and μ p \mu_p μ p , and FaSHPS rejects only candidates with A ˉ q > μ p ( 1 + z α c A / N ) ≥ 2 μ p \bar{A}_q > \mu_p(1 + z_{\alpha}c_A/\sqrt{N}) \ge2\mu_p A ˉ q > μ p ( 1 + z α c A / N ) ≥ 2 μ p .
By Eq. (37) , D p q + 8 / π D_{pq} + 8/\pi D pq + 8/ π is 4 / π 4/\pi 4/ π times the product of an intensity term ( I ˉ p 2 + I ˉ q 2 ) / ( I ˉ p I ˉ q ) (\bar{I}_p^2 + \bar{I}_q^2)/(\bar{I}_p\bar{I}_q) ( I ˉ p 2 + I ˉ q 2 ) / ( I ˉ p I ˉ q ) , which is at least two, and a phase term, the mean over the interferograms, which is at least one. In the second pass, I ˉ r \bar{I}_r I ˉ r and the phases of the multilooked interferograms of the reference set take the place of I ˉ p \bar{I}_p I ˉ p and the phases of the center. The phase term is at most the mean of ( 1 + γ a 2 ( n ) ) / ( 1 − γ a 2 ( n ) ) (1 + \gamma_a^2(n))/(1 - \gamma_a^2(n)) ( 1 + γ a 2 ( n )) / ( 1 − γ a 2 ( n )) over the interferograms, and with this bound, on each stack and in each configuration, the statistic for equal intensities stays below the smallest first-pass threshold of its own table. On its own, the phase term thus cannot reject a candidate in the first pass, and in the second pass it rejects fewer than 0.01 % of the candidates in the sampled windows. The intensity term on its own, with equal phases, accounts for 89 to 91 % of the rejections of the second pass on the real stacks and 84 to 87 % in the simulated scenes, and the rest needs both terms. On Visp, the contrast rises with the PCI to a median of 5.0 to 7.5 dB above 0.3, against 1.9 to 2.8 dB in the simulated scenes of TSX-44, and CMP accepts 29 to 35 % of the unmasked pixels of the search window at these centers.
On the Sentinel-1 test areas, 61 to 90 % of the centers have a PCI below 0.3, against 61 % (S1-79) and 55 % (S1-20) in the simulated scenes. Below 0.3, the mean amplitudes spread about as much as in simulated clutter with strong texture or more. In each of the two bins below 0.3, FaSHPS accepts 8.1 to 29 percentage points less than in the matching simulated configuration, whereas above 0.3 the two differ by at most 11 points. With the bins weighted by their shares of the stack, at least 80 % of the difference within the bins arises below 0.3 on every stack.
Runtime
This section describes the runtime measurement of § 3.1.4 and gives the time of each step (Table S17 ). All timed methods are written as single-threaded C routines. The routines are compiled with GCC 11.4 at optimization level -O3 for the processor of the test machine, an Intel Xeon E5-2697 v3 (-march=native), which enables automatic vectorization of the inner loops, and with -ffast-math without the finite-math assumption. Only the score of the truncation step is compiled without fast-math, so that its values are bitwise identical to those of the simulations. The timed center pixels lie on a regular grid in several regions of each stack, without the masked pixels. At every center pixel the methods are run one after the other in a random order, and this is repeated in five passes, so that changes of the machine state during the
measurement affect all methods alike. The NSPC map of the mask is timed per image pixel over the same regions. We report times relative to the full window of the same scene, because absolute times depend on the machine. The full window computes the covariance over the ensemble of each center directly. Running sums would make it cheaper, but they are not possible for adaptive ensembles. The same measurement, with the same code, ran on an Intel Core i7-9700 and an AMD Ryzen 7 4800H (§ 3.1.4 ). Two complete runs of the measurement agree to within 0.068 in every ratio, and to within 0.028 except for the full window with the truncation step on Visp. Every comparison of § 3.1.4 between a method with ANS and the same SHP selection alone, the same SHP selection with the truncation step alone, or the full window has the same sign in both runs.
Table 4 gives the total time of each method in this measurement divided by the number of centers. The time of each step comes from a separate measurement, and Table S17 gives its medians over 150 center pixels per stack, the best of three runs each, so that the steps need not add up to the times of Table 4.
Timed step by step (Table S17 ), computing S p q S_{pq} S pq for all pixels q q q of the search window costs 2.8 to 7.5 % of one sample covariance over the full window, of the order of the ratio 2 / ( N + 1 ) 2/(N+1) 2/ ( N + 1 ) of the numbers of multiplications and additions per pixel. Ordering and truncation step together cost 0.43 to 1.24 times as much as one sample covariance, depending on the configuration, and the PCI 0.0008 to 0.0055 of one sample covariance per image pixel. The part of the time of the truncation step (§ 2.2.6 ) that grows with R R R , i.e., the score and its noise, is estimated from the same step timed with R = 1 R = 1 R = 1 , 2, 4, 8 and 16 reference acquisitions on the same 150 center pixels per stack. The time of the step is linear in R R R to within 2.2 %. The share of this part at R = 8 R = 8 R = 8 is eight times the slope of that line divided by the time at R = 8 R = 8 R = 8 . It is 63 to 75 % on the test machine (Visp, slow-deformation test area and landslide substack), so that most of the time of the truncation step is proportional to R R R .
The truncation step costs relatively more on Visp than on the Sentinel-1 test areas (Table 4). The cost of its score grows with R N RN R N per member and that of the coherence matrix with N 2 N^2 N 2 , so that the step costs relatively more with the 44 acquisitions of Visp than with 79 (Table S17 ). On Visp the ensembles are also less often truncated than on the Sentinel-1 test areas, so that less of the coherence matrix is saved.
Relative to the processing with the SHP selection alone, ANS adds at most 0.09 times the processing time of the full window on Visp and the Maiun test area and saves time on the slow-deformation and long-term coherent test areas, for the full window, FaSHPS and CMP (Table 4 ). At the timed center pixels that pass the gate, the PCI-scaled window holds on average 79 % (Visp) and 69 to 81 % (Sentinel-1) of the pixels of the search window, which reduces the time of the SHP selection, the truncation step and the coherence matrix.
Measures without ground truth
This section gives the protocol and the remaining numbers of § 3.1.5 . Both measures, the split-stack agreement and the predicted phase standard deviation, were evaluated on the first 40 random scenes of each batch of evaluation scenes and on control scenes. The control scenes are homogeneous clutter with a planar phase ramp of at most one fringe per window diameter, 80 per configuration with the exponential coherence model (Eq. (27) ) and as many with a coherence equal for every pair. In the controls with the exponential model, the split-stack agreement prefers CMP+ANS to the full window with ANS in 48 to 60 % of the scenes, whereas the phase linking error of the split-stack history does so in 34 to 60 %. In the controls with a coherence equal for every pair, the two shares are 38 to 55 % and 36 to 49 %, respectively. Where neither SHP selection should be preferred, neither measure thus gives a clear preference. Between an SHP selection alone and the same SHP selection with ANS, the split-stack agreement ranks as the phase linking error of the split-stack history does in 78 to 88 % of the random scenes in which these errors differ by more than 5 % (TSX-44 and S1-79, § 3.1.5 ). In the random scenes where the split-stack agreement of an SHP selection alone and of the same SHP selection with ANS differs by more than 5 %, the measure orders the two as the phase linking error of the split-stack history does in 84 to 93 % of the scenes (three configurations). Here the scenes are selected by their difference in agreement, and in § 3.1.5 by their difference in phase linking error. The split-stack agreement prefers ANS more often than the phase linking error of the split-stack history does in seven of the nine pairs of SHP selection and configuration, and in 74 to 82 % of the scenes against 67 to 78 % over the nine pairs.
The 16 regions of the real stacks are four on Visp and three on each of the Maiun, long-term coherent, slow-deformation and fast-deformation test areas of Sentinel-1 with 79 acquisitions. There the two measures repeat the tendencies of the simulation. The split-stack agreement ranks an SHP selection with ANS first in 13 regions, and the predicted standard deviation ranks the full window alone first in all of them. Between an SHP selection alone and the same SHP selection with ANS, the split-stack agreement is higher with ANS in 43 of the 48 pairs of SHP selection and region.
Further methods and options
This section reports further methods and options run under the protocol of § 2.3 , to which § 2.2 and § 3.2 refer. In this section, a value or a range refers to all centers of the evaluation scenes, the full window, FaSHPS and CMP, and the three configurations TSX-44, S1-79 and S1-20, unless the sentence or the parentheses after it name others. A range gives the smallest and the largest value over them.
Reference acquisitions of the score Table S18 gives the error of FaSHPS+ANS with other sets of reference acquisitions in the score (Eq. (23) ). More references add pairs to the score and lower the error slightly, at a proportionally higher cost. In homogeneous clutter of coherence 0.5, σ Δ α \sigma_{\Delta\alpha} σ Δ α with 16 references is 0.72 to 0.81 times that with eight, and with all acquisitions 0.47 to 0.79 times, so that the margin ζ \zeta ζ falls with it. These ratios are medians over the centers of the scenes of the paragraph on the noise of the gain below (three configurations). The error over all centers falls by at most 1.8 %. The cost of the score, however, is proportional to ∣ I ∣ |\mathcal{I}| ∣ I ∣ , which grows from 8 N − 36 8N - 36 8 N − 36 to 16 N − 136 16N - 136 16 N − 136 and N ( N − 1 ) / 2 N(N - 1)/2 N ( N − 1 ) /2 . With a single reference, the first or the middle acquisition, leaving out this reference in the jackknife (Eq. (39) ) leaves no pair, and its value Δ α ( s ) \Delta\alpha_{(s)} Δ α ( s ) is taken as Δ α \Delta\alpha Δ α . The error over all centers then rises by 3.7 to 8.3 %, and the outcome depends on the choice of the reference. The two single references return the same ensemble at 29 to 39 % of the centers at which the gate applies the truncation step.
Table S18. Options of the truncation step within the PCI-scaled window and the gate, applied after FaSHPS: phase linking error relative to the oracle, weighted geometric mean over all centers of the evaluation scenes and over the distributed centers with two or more fringes per window diameter.
Fringe frequency compensation The truncation step reduces bias at the expense of looks. Where fringes are dense and coherence is low, both a low bias and many looks are required. A known remedy for single interferograms is to estimate the local fringe frequency and to subtract, before averaging, the local fringe pattern as a phase plane that is zero at the center, as in the slope-compensated filter of Trouvé et al. [19 ] . We examined a simple version with FaSHPS and the PCI-scaled window, applied at the centers that pass the gate. For each interferogram of two consecutive acquisitions, a local wavenumber is estimated along each image axis. To this end, the interferogram is multiplied by its conjugate, shifted along the axis by more than the speckle correlation length. The product is averaged over a box and over five consecutive interferograms, and the phase of this average, divided by the shift, is the local wavenumber. Where the magnitude of the average lies below a threshold, so that its phase is mostly noise, the wavenumber is set to zero. Summed over the interferograms from the first acquisition, the wavenumbers give the local phase gradient k p ( n ) k_p(n) k p ( n ) of acquisition n n n at center p p p , relative to the first acquisition. The data in the PCI-scaled window of p p p is multiplied by exp { − i k p ( n ) ⋅ ( q − p ) } \exp\{-i k_p(n)\cdot(q-p)\} exp { − i k p ( n ) ⋅ ( q − p )} , which leaves the phase of the center unchanged. Since an inaccurate ramp adds fringes instead of removing them, the SHP selection and the truncation step are applied to the compensated and to the uncompensated window, and the compensated ensemble is retained only if its score (Eq. (23) ) exceeds that of the uncompensated one by more than 0.5 / N 0.5/N 0.5/ N .
With this safeguard, the compensated ensemble was kept at 2.5 to 5.1 % of all centers. Relative to FaSHPS+ANS without compensation, the error over all centers changed by −0.6 to −0.1 % and that at two or more fringes per window diameter by −5.0 to −1.3 % (Table S18 , three configurations). Without the safeguard, the error over all centers was 1.27 to 1.69 times that without compensation. This ratio was largest at c = 0.35 c=0.35 c = 0.35 , at 2.2 to 4.3, where the magnitude of the average lies above the threshold but its phase is still noisy, and the sum of the wavenumbers over the stack drifts. Since the compensation with the safeguard changes the error over all centers by less than 1 % and depends on the safeguard, it is not used.
Local polynomial fit The local polynomial approximation with the intersection of confidence intervals ([8 ] ) fits a phase plane, a phase and two slopes, to the phasors of one interferogram over nested windows and keeps the largest window up to which the confidence intervals of all nested windows have a point in common. It returns a phase per interferogram and no ensemble. We applied it to the N ( N − 1 ) / 2 N(N-1)/2 N ( N − 1 ) /2 interferograms of every center of the first batch of evaluation scenes and linked the resulting matrix with the same phase linking (Table S19 ). The window fractions ran from 0.1 to 1 of the search window, and the threshold factor of the confidence intervals was fixed at 2 without tuning. On the same centers, its error was higher than those of FaSHPS+ANS and FaSHPS+T over all centers, in each coherence class of Table S19 and at two or more fringes per window diameter, in the three configurations. Over all centers it was 1.77 to 1.92 times the oracle error, against 1.22 to 1.32 for FaSHPS+ANS, and it was highest at c = 0.2 c = 0.2 c = 0.2 , at 3.05 to 4.04 times the oracle error. Its excess over FaSHPS+T was smallest at c = 0.95 c = 0.95 c = 0.95 in TSX-44, 8.4 %, so that it has more error than FaSHPS+T in every coherence class.
Table S19. Local polynomial fit with the intersection of confidence intervals, applied to every interferogram, compared with the full window and FaSHPS alone, with the truncation step (T) and with the PCI-scaled window and the gated truncation step (ANS): phase linking error relative to the oracle, weighted geometric mean over the centers of the first batch of evaluation scenes (all centers, distributed centers by coherence, and distributed centers with two or more fringes per window diameter).
Variants of the truncation step and of FaSHPS Table S20 gives the error of the full window, FaSHPS and CMP with variants of ANS, relative to the same SHP selection with ANS at the selected setting, on the centers of the evaluation scenes. The noise multiplier λ \lambda λ limits truncation where the core carries almost no information about the true phase (§ 2.2.6 ). Where such a core happens to disagree with the true phase, smaller subsets reach a higher score in the truncation step (Eq. (23) }), because their noisier multilooked phases bring the mean cosine closer to zero. A truncation there changes the phase linking error little, since the phases are noise either way. With $\lambda = 0$, 1, 2, 4 and 5 and the selected $\kappa$, the error over all centers changes by at most 0.4 %, and on the tuning scenes the criterion changes by less than 0.1 % for every value of $\lambda$ in Table S2`. Ordering the members by their SPC with the center instead of the ellipse order raises the error to 1.23 to 1.52 times that at the selected setting, and growing the ensemble in this order up to the candidate size of largest Fisher information raises it to 1.14 to 1.28 times. The variant of FaSHPS+ANS of Section 6 that also accepts candidates by their SPC has 1.02 to 1.04 times the error of FaSHPS+ANS.
Distribution tests Table S21 gives the error with the Kolmogorov–Smirnov test of SqueeSAR (Ferretti et al., [3 ] ), Kuiper’s test (Kuiper, [9 ] ), the Anderson–Darling test (Scholz and Stephens, [16 ] ) and the generalized likelihood ratio test (Parizzi and Brcic, [14 ] ) in place of the amplitude test of FaSHPS, at a significance level of 1 %. Each test is used with a threshold that depends on η \eta η , calibrated in homogeneous clutter like that of FaSHPS, and with the nominal threshold for independent samples. With the threshold
Fixed constants Table S22 gives the error of the full window and FaSHPS with ANS when one of its fixed constants (Table 2 and § 2.2 ) is changed, relative to the same method with ANS at those values, on the centers of the evaluation scenes. Over all centers, the error changes by at most 2.1 %, most with R = 4 R = 4 R = 4 reference acquisitions. The sweep also covers the tuned bound ρ min \rho_{\min} ρ m i n . Relative to ρ min = 0.025 \rho_{\min} = 0.025 ρ m i n = 0.025 , the error with ρ min = 0.02 \rho_{\min} = 0.02 ρ m i n = 0.02 changes by at most 0.2 % over all centers and is at most 1.5 % higher at point scatterers, and with ρ min = 0.03 \rho_{\min} = 0.03 ρ m i n = 0.03 it is 0.6 to 0.9 % higher over all centers and 0.4 to 5.4 % lower at point scatterers, the trade-off of limitation (viii) of § 3.2 .
Bias of the score If the core pixels share the true phase history θ p \theta_p θ p of the center, the CCG model (§ 2.1.4 ) gives E [ v ˉ C ( n , r ) ] = g ˉ C ( n , r ) e i ( θ p ( n ) − θ p ( r ) ) \mathbb{E}[\bar{v}_C(n,r)] = \bar{g}_C(n,r)e^{i(\theta_p(n)-\theta_p(r))} E [ v ˉ C ( n , r )] = g ˉ C ( n , r ) e i ( θ p ( n ) − θ p ( r )) for the multilooked interferogram of the core (Eq. (22) ), where g ˉ C ( n , r ) \bar{g}_C(n,r) g ˉ C ( n , r ) is the mean of g ( ∣ γ n r p ∣ ) g(\lvert\gamma_{nr}^{p}\rvert) g (∣ γ n r p ∣) over the core pixels p p p . If their speckle is also independent of that of the scored pixels, which the guard set provides, the expectation of the score (Eq. (23) ) over the speckle of the core, with E k E_k E k fixed, is therefore
E [ a k ∣ E k ] = 1 ∣ I ∣ ∑ ( n , r ) ∈ I g ˉ C ( n , r ) cos ε k ( n , r ) , (38) \mathbb{E}[a_k\mid E_k]=\frac{1}{\lvert\mathcal{I}\rvert}\sum_{(n,r)\in\mathcal{I}}\bar{g}_C(n,r)\cos\varepsilon_k(n,r), \tag*{(38)} E [ a k ∣ E k ] = ∣ I ∣ 1 ( n , r ) ∈ I ∑ g ˉ C ( n , r ) cos ε k ( n , r ) , (38)
where ε k ( n , r ) = arg e ^ k ( n , r ) − θ p ( n ) + θ p ( r ) \varepsilon_k(n,r)=\arg\hat{e}_k(n,r)-\theta_p(n)+\theta_p(r) ε k ( n , r ) = arg e ^ k ( n , r ) − θ p ( n ) + θ p ( r ) . The score is thus a mean agreement, weighted with the coherence, of the multilooked phases with the true phase of the center, in the form of Eq. (8) . Since the core contains the center, pixels chosen by their agreement with the phases of the center could, however, score higher than their true phase justifies. CMP chooses them in its first pass by the single-look phase difference to the center, and the ellipse order uses the SPC of the center only through the shape of a moment matrix. Table S23 compares the score of the truncation step (Eq. (23) ) with its expectation over the speckle of the core (Eq. (38) ), computed from the true phase and coherence of every core pixel with the ensemble fixed. It uses the ensembles of the full window, FaSHPS and CMP in the search window without the PCI-scaled window, in the ellipse order and in the distance order, on the first 100 scenes of each batch of evaluation scenes, at the centers without a point scatterer in the core. Averaged over the candidate sizes, the difference is − 0.001 -0.001 − 0.001 to 0.002 0.002 0.002 , against mean expected scores of 0.26 0.26 0.26 to 0.34 0.34 0.34 in the ellipse order, and within each coherence class the difference for FaSHPS or CMP minus that for the full window is − 0.003 -0.003 − 0.003 to 0.000 0.000 0.000 . With the expected scores in place of the observed ones, the truncation changes at up to 22% of these centers, for the full window as for FaSHPS and CMP. Since the full window selects no member by its phase, this share reflects the noise of the core. In the simulations, the pairwise SHP selections and the ellipse order thus add no positive bias.
Table S23. Bias of the score of the truncation step on the first 100 scenes of each batch of evaluation scenes. Observed score (Eq. (23) ) minus its expectation from the truth (Eq. (38) ), averaged over the candidate sizes and as a weighted mean over the centers without a point scatterer in the core (b b b ), the share of these centers whose truncation changes when the expected scores decide (chg.) and the share truncated with the observed scores (tr.), for the ensembles of the search window in the ellipse order and in the distance order.
Noise of the gain Independent terms would give the gain Δ a \Delta a Δ a of § 2.2.6 the standard deviation ∣ I ∣ − 1 ( ∑ ( n , r ) ∈ I Δ y ( n , r ) 2 ) 1 / 2 \left\lvert\mathcal{I}\right\rvert^{-1}\left(\sum_{(n,r)\in\mathcal{I}}\Delta y(n,r)^2\right)^{1/2} ∣ I ∣ − 1 ( ∑ ( n , r ) ∈ I Δ y ( n , r ) 2 ) 1/2 . Where the core carries information about the true phase, terms that share an acquisition are correlated, because a deviation of the phase of that acquisition in the core or in the ensemble shifts all of them. The jackknife over acquisitions (Efron and Stein, [1 ] , Eqs. (1.1) to (1.3)) leaves out each acquisition in turn. With I s ⊂ I \mathcal{I}_s \subset\mathcal{I} I s ⊂ I the pairs that contain acquisition s s s , it gives
Δ a ( s ) = 1 ∣ I ∣ − ∣ I s ∣ ∑ ( n , r ) ∈ I ∖ I s Δ y ( n , r ) , σ Δ a = ( N − 1 N ∑ s = 1 N ( Δ a ( s ) − Δ a ‾ ) 2 ) 1 / 2 , (39) \Delta a_{(s)}=\frac{1}{\lvert\mathcal{I}\rvert-\lvert\mathcal{I}_s\rvert}\sum_{(n,r)\in\mathcal{I}\setminus\mathcal{I}_s}\Delta y(n,r),\qquad\sigma_{\Delta a}=\left(\frac{N-1}{N}\sum_{s=1}^{N}\left(\Delta a_{(s)}-\overline{\Delta a}\right)^2\right)^{1/2}, \tag*{(39)} Δ a ( s ) = ∣ I ∣ − ∣ I s ∣ 1 ( n , r ) ∈ I ∖ I s ∑ Δ y ( n , r ) , σ Δ a = ( N N − 1 s = 1 ∑ N ( Δ a ( s ) − Δ a ) 2 ) 1/2 , (39)
with Δ a ‾ \overline{\Delta a} Δ a the mean of the N N N values Δ a ( s ) \Delta a_{(s)} Δ a ( s ) . The jackknife treats the acquisitions as exchangeable, although a reference acquisition belongs to N − 1 N-1 N − 1 pairs and any other to R R R , and σ Δ a \sigma_{\Delta a} σ Δ a rests on N N N values, 20 in the shortest stack. The following values are medians over the centers at which a smaller subset has the highest score, since elsewhere Δ a = 0 \Delta a=0 Δ a = 0 , among 120 random centers of one scene per configuration with the full window in the ellipse order as the ensemble. In homogeneous clutter of coherence 0.2, where no truncation is justified, the jackknife estimate σ Δ a \sigma_{\Delta a} σ Δ a is 1.52 to 1.65 times the standard deviation for independent terms, 1.86 to 2.36 times at a coherence of 0.5 and 2.53 to 4.17 times at 0.95. The terms are thus correlated, and the standard deviation for independent terms would underestimate the noise of the gain. Table S24 compares σ Δ a \sigma_{\Delta a} σ Δ a with the actual noise of Δ a \Delta a Δ a , i.e., its standard deviation over independent centers in homogeneous clutter, for the smallest, the sixth and the largest candidate size L K − 1 L_{K-1} L K − 1 of the distance order. The root mean square of σ Δ a \sigma_{\Delta a} σ Δ a is 1.40 to 1.51 times that standard deviation at c = 0 c=0 c = 0 , for every candidate size. At c ≥ 0.5 c\ge0.5 c ≥ 0.5 it is 0.60 to 0.80 times for the smallest candidate, lower than for the larger candidates at the same coherence (0.79 to 1.09, the three configurations and S1-79 on the native range grid, S1-79n). Since κ \kappa κ is tuned, a common factor would not matter, but the factor varies with the coherence and the candidate size. The margin κ σ Δ a \kappa\sigma_{\Delta a} κ σ Δ a is thus wider than κ \kappa κ times the noise at c = 0 c=0 c = 0 , where the gate skips the truncation step at most distributed centers, and narrower for the smallest candidate in coherent clutter.
Table S24. Jackknife noise of the gain in homogeneous clutter of coherence c c c : root mean square of σ Δ a \sigma_{\Delta a} σ Δ a over the centers divided by the standard deviation of Δ a \Delta a Δ a over the same centers, for three candidate sizes of the distance order. A calibrated estimate gives one. The three sizes have 25, 189 and 4865 members in TSX-44, 25, 189 and 2162 in S1-79 and S1-20, and 25, 189 and 961 in S1-79n, S1-79 on the native range grid (Table S25 ). Every center has its own search window, the windows do not overlap, and the scenes are independent (1,800 centers in TSX-44, 1,800 centers in S1-79, 1,800 centers in S1-20, 1,200 centers in S1-79n per coherence).
Native Sentinel-1 sampling Table S25 repeats the main comparison of S1-79 on the native range grid, with every second range sample of the same scenes and with calibration scenes of its own (S1-79n). On the native grid, ANS lowers the error about as much as on the oversampled grid of S1-79, over all centers by 21 to 22 % relative to the SHP selection alone and at one fringe or more per window diameter by 47 % (three SHP selections).
Table S25. S1-79 on the oversampled grid and on the native grid of the same scenes (S1-79n, every second range sample kept, with its own calibration scenes). Phase linking error relative to the oracle of each configuration, groups, variants and columns as in Table 3 . The S1-79 block repeats Table 3 .