When the Comparison Is the Problem: Spatial Resolution and Validation Bias in InSAR-Derived Coastal Subsidence Assessments Along the U.S. Gulf Coast

Comment #316 Manoochehr Shirzaei @ 2026-07-11 22:51

Response to Reviewer#1 Comment:

Re: Comment on Li et al. (2026), “When the Comparison Is the Problem: Spatial Resolution and Validation Bias in InSAR-Derived Coastal Subsidence Assessments Along the U.S. Gulf Coast”
We thank the reviewer for a thoughtful and substantive comment, and we agree that comparison and debate are how this field improves. We welcome it. But welcoming the debate is not the same as accepting the terms on which it has been framed here, and we want to address each point directly.
On the inference from disagreement to “not bare-earth VLM”
The reviewer is right that vegetation cover challenges C-band interferometry, and right that L-band systems penetrate canopy more readily. Neither point is in dispute. However, there is also a factual problem with how the C-band argument is applied here. The comment treats Ohenhen et al. (2024) as a single-sensor (i.e., C-band) product and bases its penetration-depth argument on that premise. It is not a single-sensor product. Ohenhen et al. combine C-band (Sentinel-1 for 2015-2020) and L-band (ALOS for 2007-2011) observations to reduce the canopy decorrelation effect that the reviewer describes. The paragraph's argument requires Ohenhen et al. to inherit Sentinel-1's vegetation penetration limits outright, and it does not. The premise on which the rest of that paragraph rests does not hold as stated.
The reviewer's central move is this: if land cover and coherence affect the VLM estimate, then the products are not measuring bare-earth motion but “some property of the vegetation cover.” We do not think this follows. Vegetation-driven decorrelation degrades the precision of a phase-based measurement and, in some cases, biases it, which affects how well the ground motion is estimated rather than what physical quantity is being estimated. The same pattern shows up throughout remote sensing: optical retrievals vary in accuracy by land cover too, and no one concludes from that variation that the sensor has stopped measuring the thing it was built to measure.
It is also worth stating plainly that Ohenhen et al.'s coherence-based elite pixel selection is specifically designed to address this problem. It excludes the pixels most vulnerable to vegetation-driven phase contamination and retains high-confidence estimates at native resolution. Li et al.'s response to the same underlying issue, aggregating from 50 m to 1 km before comparison, does the opposite. It blends low-coherence, contamination-prone pixels with clean ones and claims the result is more robust. That is not a solution to vegetation contamination; it is a way of concealing it inside a coarser average. If the reviewer's concern about land-cover-dependent bias is taken seriously, it argues against Li et al.'s aggregation approach, not against the higher-resolution product from which it was aggregated.
On variable uncertainty by coherence
We agree with the underlying principle. Letting uncertainty scale with coherence, or with some other vegetation-related property, is good practice and a direction the field should move toward in general. We are glad to say so.
But this is not a gap in Ohenhen et al. (2024). It is worth pointing out that the product already does this. It is built on a stochastic estimation approach that combines L-band and C-band observations with GNSS, weighting each input by its own uncertainty, with coherence directly entering that weighting. The output is a full 3D displacement field with associated uncertainties propagated through the same framework. Thus, the uncertainties already vary pixel-by-pixel with land-cover and phase-observation noise. This is not a proposal for future work in our case. It is the estimation framework already in use.
So if the reviewer's suggestion is worth adopting, it should be adopted by using the pixel-level uncertainty structure as it already exists in Ohenhen et al., rather than discarding it. Li et al.'s approach does the opposite: it aggregates to 1 km and applies a single fixed 5 mm/yr threshold across the entire domain, which erases exactly the coherence-dependent uncertainty information the reviewer is asking the field to preserve. The tool the reviewer wants already exists in the finer-resolution product provided by Ohenhen et al. It was the coarsening step, not the original dataset, that threw it away. That is a separate point from whether the epoch- and resolution-mismatched comparison in Li et al. was fair to begin with, and we do not think either one substitutes for the other, but on this specific question, the fix the reviewer is calling for is not something still owed. It is something already built, and already lost in translation on the other side of the comparison.
On accuracy versus precision
It is worth pausing to separate two concepts that the comments repeatedly run together: accuracy and precision. Accuracy is closeness to the true value. Precision is the spread of an estimate around whatever value it is centered on. The two are independent properties of an estimator, not two words for the same thing. A result can be accurate but imprecise, or precise but inaccurate.
In our framework, these are assessed separately, against separate references. Comparison with GNSS, treated as the ground truth, serves as a check on accuracy: it asks how close the InSAR-derived VLM is to an independent measurement of the true motion at that location. The per-pixel standard deviation produced by propagating phase noise and coherence through the stochastic estimation framework serves as a check on precision: it asks how tightly constrained a given estimate is, regardless of whether GNSS happens to agree with it. Ohenhen et al. report both, separately, because they answer different questions.
The comment does not maintain this distinction, and neither does Li et al. Disagreement between two independent InSAR products in vegetated terrain is treated as evidence that the product cannot be trusted, implicitly, an accuracy failure. But two unbiased estimators can disagree with each other simply because one has larger variance than the other under certain conditions; disagreement by itself does not say which, if either, is biased. The reviewer's own suggestion that “the two results in vegetated environments are actually equivalent within realistic uncertainties” is close to the right question. But it is a question about precision, not accuracy, and it is one our pixel-level uncertainty estimates were already built to answer.
This distinction also exposes a problem with spatial aggregation as a proposed fix. Averaging pixels together narrows the spread of the aggregate almost by construction, regardless of whether the underlying pixels are individually biased. That narrowing looks like an improvement, but it is a precision effect, not an accuracy result. If some aggregated pixels exhibit vegetation-driven bias, averaging them with clean pixels can degrade accuracy even as the aggregate appears tighter and more self-consistent. A result that looks more precise after coarsening is not, on that basis alone, more accurate.
On GNSS validation sites and scattering characteristics
This is the strongest point in the comment, and we want to engage with it directly rather than dismiss it. The reviewer's observation that GNSS monuments are cleared of vegetation, carry metallic superstructures, and are often fenced is a legitimate concern about what GNSS-based validation actually tests. A site built for a clean sky view does not necessarily reproduce the radar scattering environment of the surrounding marsh, and comparisons that rely on GNSS agreement as a general accuracy statement should be read with that in mind.
That said, this concern does not support the conclusion the reviewer draws from it, for two reasons. First, it is symmetric. If GNSS validation sites are systematically biased toward favorable scattering conditions, that bias applies to every InSAR product validated against them, including Wang et al., Ohenhen et al., and any future reprocessing of Li et al.'s approach. It is a caution about the limits of GNSS as ground truth in vegetated terrain, in general. It is not evidence that favors a coarsened, spatially aggregated product over a native-resolution one. Second, and more directly, our original comment's objection to Li et al. does not rest on GNSS validation at all. It rests on the observation that the two InSAR products being compared cover different epochs in a system that Li et al. themselves describe as temporally nonlinear, and that they were compared at resolutions that differ by roughly two orders of magnitude. Whatever the GNSS network's siting characteristics turn out to be, they have no bearing on whether that comparison was internally fair.
Summary
We welcome the reviewer's invitation to debate, and we think debate is exactly what this exchange has been. But debate has to compare like with like. Our objection to Li et al. was never that disagreement between InSAR products is impossible or embarrassing; it was that this particular comparison set two products against each other across different epochs and different resolutions, and then attributed the resulting discrepancy to the finer of the two. Nothing in this comment changes that. Underneath several of the specific points, we think the comment also conflates accuracy with precision: disagreement between products in vegetated terrain is read as an accuracy failure, when it is at least as plausibly a precision effect that our pixel-level uncertainty framework was already built to capture, and that spatial aggregation obscures rather than resolves. It is also worth noting that the comment's opening premise that Ohenhen et al. share Sentinel-1's C-band penetration limits is factually incorrect, since Ohenhen et al. is a C-band/L-band fusion product built in part to address exactly that limitation. The reviewer's strongest point, on GNSS siting, is a fair caution about validation in general. But it is not a defense of the specific comparison our comment addressed.

__________________________________________________________________________________________________
Reviewer#1 comment:

RC1: 'Comment on egusphere-2026-3372', Timothy H. Dixon, 08 Jul 2026 reply
It is well known that densely vegetated terrain is a challenging environment for SAR interferometry. This is especially true for shorter wavelength C-band radars such as Sentinel-1, compared to longer wavelength L-band radars such as ALOS-2 and NISAR, because C-band radars are less likely to penetrate vegetation cover and hence have difficulty recording changes in ‘bare earth’ elevation (vertical land motion, VLM). Two recent publications (Wang et al., 2024; Ohenhen et al., 2024) attempt to measure VLM in low-lying, often vegetated coastal terrain, an important target because of the likelihood of subsidence and consequent flood hazard. The authors apply sophisticated approaches to address the loss of interferometric phase coherence associated with vegetation cover. Wang et al. (2024) use an eigenvalue decomposition approach for phase reconstruction, whereas Ohenhen et al. (2024) use a filtering approach to eliminate low-coherence pixels. The two techniques give similar results in urban areas where vegetation is limited and strong radar scatterers are numerous. However, in densely vegetated coastal terrain, the two techniques can give different results. Li et al (2026) pointed this out, and suggest that in such coastal terrains, it would be prudent to assume a threshold confidence limit of 5 mm/yr for InSAR until the source of disagreement between the various approaches is better understood.

In their comment on Li et al, (2026) Shirzaei et al. suggest that the comparison of the two results is not valid because the spatial resolution of the two techniques being compared differs. Li et al. average across different land cover types and coherence conditions to compare equivalent size pixels. Shirzaei et al. suggest this approach is not appropriate because of land cover and coherence differences. However, if landcover type and coherence have a big impact on the estimate of VLM, this implies that it is not bare earth VLM that is being measured, but rather some property of the vegetation cover.

In their abstract, Shirzaei et al. state that inter-product disagreement should not be conflated with measurement failure. Perhaps, but until that disagreement is better understood, it should certainly give our community pause for reflection. At a minimum, I suggest we consider putting larger uncertainty bounds on VLM estimates derived from C-band radar data in heavily vegetated terrain. Perhaps the apparent disagreement between the Wang et al. and Ohenhen et al. results in coastal vegetated environments simply reflects underestimated uncertainties – perhaps the two results in vegetated environments are actually equivalent within realistic uncertainties. Let’s not assign a single, overall uncertainty to our VLM estimates, but rather allow uncertainties to vary with coherence or some other vegetation-related property.

Both Wang et al. and Ohenhen et al. calibrate their InSAR results and assess accuracy by comparing with VLM data from GNSS stations. In my opinion this gives an overly optimistic view of performance. Here’s why. The two InSAR results agree well in urban areas, suggesting better performance (accuracy) in settings with strong radar reflectors. GNSS stations are distributed widely, in both urban and rural vegetated areas. However, GNSS sites in rural areas are not typical of the radar scattering characteristics of their surroundings. The sites are cleared of vegetation (otherwise the GNSS antenna would lack good sky view), and typically have metallic superstructure for the GNSS antenna, ancillary electronics, communications, and power. The GNSS sites are often surrounded by a chain link fence. In effect, the sites have the radar scattering characteristics of an urban location. So it is not surprising there is good agreement between the InSAR and GNSS estimates of VLM.

Comparison and debate are healthy in science. Let’s welcome Li et al.’s (2026) comparison of two published results and see how we can improve the science.

References

Li et al. (2026) Earth Obs, 1, 1-13

Ohenhen et al. (2024) Nature, 627, 108-115.

Wang et al. (2024) J. Geophys. Res. – Earth Surface, 129

Comment #308 Manoochehr Shirzaei @ 2026-06-10 19:08

This article is now submitted to Earth Observation journal as comment/reply. please consider participating in online conversation and review.
Thanks
Manoo