The Ground Moved 30 mm. Or Did It? How to Tell Real InSAR Subsidence From Processing Error

The Ground Moved 30 mm. Or Did It? How to Tell Real InSAR Subsidence From Processing Error

Your satellite map shows 30mm of ground movement. Before you go blathering on about the foundation, prove the ground has actually moved.

That sounds a bit harsh, so let me put it in a real situation. A client sends you a deformation map. There are red points along a road, or around a dam, or under a block of buildings. There's a time series that's going down. The one question is "Is this real?" You're the surveyor, so that's the question that's on you.

Here's the bit that should have you pausing for a moment. Sentinel-1 is a C-band satellite, and one full fringe in an interferogram equates to about 2.8cm of change along the line of sight (that's half of its 5.55cm wavelength). A single phase unwrapping slip adds or subtracts one full cycle. Simple maths tells you that's about 28mm. Your headline figure is essentially the size of one processing error.

This is even more important now than it was two years ago. The NASA NISAR L-band data has been available via the ASF DAAC since July 2026. OPERA has now got free 30m displacement products over North America, and Europe has had the Copernicus ground motion service for some time. More and more people are now giving you InSAR maps without ever having actually seen an error budget. If you read our earlier post on whether a change detection map actually shows change, then this one is the radar phase equivalent of that same question.

So here's how you read one of these maps without fooling yourself.

What an InSAR number actually measures

It is relative, not absolute

An InSAR displacement value is not a surveyed height change. It's a change in phase between acquisitions which has been converted into a distance. It only makes sense in relation to two things: a reference point (which is assumed to be stable) and a start date (which is set to zero).

Move the reference point and the entire map shifts. That's not a minor issue. Reference point selection is listed in the literature on time series InSAR as one of the factors that limits the size of the signal you can detect. If your "stable" point is actually on a slowly subsiding fill embankment, then everything else around it is going to look like it's rising.

It is measured along the line of sight

The satellite looks down at the ground at an angle. What it measures is the movement of the ground towards or away from the satellite along that slanted line of sight, not the movement of the ground straight up or down (i.e. vertically). We'll come back to this as it's the very point at which most of the wrong conclusions are made.

Know which product you are holding

The Copernicus EGMS Basic product is referenced to a local point, so you can't compare movement between different processing units. The Ortho product is a completely different product: it takes the ascending and descending passes and combines them into vertical and east-west layers on a 100m grid, referenced to GNSS. Same service, very different questions you can actually answer with it. If the screenshot your client has sent doesn't tell you which product it is, then that's your first question.

Reference frames are a problem in drone work as well (see why repeat drone surveys don't line up even when nothing has changed). The same principle applies here.

Where false movement comes from

Tropospheric delay and unwrapping errors are the two dominant error sources in the multi-interferogram analysis. They are by no means the only error sources.

Atmosphere

Water vapour affects the speed at which the radar signal travels, and the interferogram cannot distinguish this from the ground moving closer to or further from the radar. Some of it is seasonal, some of it is random, and weather model corrections help but don't eliminate all of it. Residuals remain, particularly at spatial scales smaller than the weather model grid. A bump that is present in one date pair and gone in the next is therefore suspect.

Unwrapping errors

Phase is only known modulo one cycle. To get a continuous displacement, the software has to "unwrap" it, and at times it gets the count wrong by one or more cycles. That gives rise to a jump which looks very real in the time series. That's the mechanism behind the 28mm figure at the beginning. There are methods to detect and correct for these errors, but they can struggle when a significant proportion of the interferograms are affected.

Vegetation and soil moisture

This one catches people out. DLR research describes a systematic short-term phase bias linked to soil moisture and land cover which can lead to velocity biases of several mm/yr. The usual suspects are fields, forest edges and irrigated land. You're looking at wet season, dry season, and a bias that looks like a trend.

Reference point and processing choices

Even with the same satellite data, you get different numbers from different algorithms. A benchmark over Glasgow which compared four different processing approaches found they all picked up the major deformation areas, with an average spread of about 1.1mm/yr between the methods. I take that as good news and a warning in equal measure. The big real signal is there. The small detail is very much dependent on the processing.

How accurate is Sentinel-1 InSAR really?

Honestly, it all depends on what you ask it to do.

One validation study of the Sentinel-1 PSI and SBAS results found that the vertical velocity differences against GNSS had a standard deviation of 9 to 10 mm/yr and 8 mm/yr against levelling, respectively. The same paper also reported sub-cm accuracy for individual records in the time series. Both of those things can be true. They're different questions entirely (velocity over an entire area versus velocity at one point on one date).

So what does that mean for your 30mm? It means the answer is not "InSAR is unreliable" and it's not "InSAR is proven". It's a measurement with an error budget, and one study doesn't give you a universal figure. Your job is to work out what the error budget is for that site, that stack, that processing, and whether 30mm is comfortably above it. A 30mm signal with an error budget of 8-10mm is a very different proposition to a 30mm signal with no uncertainty at all stated.

Is InSAR displacement the same as vertical settlement?

No, and this is where the casual errors are most likely to occur.

The assumption hiding in the formula

To convert line-of-sight movement into vertical movement, you can simply divide by the cosine of the incidence angle. That formula is only valid if there is absolutely no horizontal movement. Researchers who looked at this found that projecting LOS to vertical in that way could differ from the correct multi-geometry result by as much as 67% of the maximum vertical rate in their example at a 33.8 degree incidence angle. Many of the published studies will use the projection approach without making any mention of the assumption.

The north-south blind spot

Even with both ascending and descending passes, you end up with two components, not three. The polar-orbiting SAR satellites are pretty much insensitive to north-south horizontal motion. The "vertical" and "east-west" layers that you see are therefore the best two-component fit under that assumption.

Where it matters most

For a broad aquifer-driven subsidence bowl, the motion is predominantly vertical, and the simple conversion is generally acceptable. For a landslide, a mine or a slope above a road, the horizontal movement can be as much as the vertical, and to simply call that "settlement" is likely to have the engineer looking at the wrong failure mechanism. You have to ask what kind of ground you're dealing with before you will even consider the term "subsidence".

Cross-checks that actually count

An independent measurement is one that doesn't have the same error sources. A second InSAR pass by the same vendor using the same reference point is not independent.

GNSS and levelling

Continuous GNSS (or repeat levelling) is the standard answer, and that's the approach that was taken in the validation studies referenced above. If there's a CORS station or a benchmark within the footprint, then check the InSAR value at that point first. Compare in the same direction (project the GNSS into the line of sight, not the other way around) so you're not making different assumptions.

Field evidence

Don't underrate cracks, tilts, door frames that won't shut, and rainfall records. In one study of the EGMS in a landslide area, the researchers actually used a building damage inventory and geomorphological mapping for spatial validation and rainfall records for temporal validation. It's very low-tech, and it works. If the satellite says 30mm and you can't find a crack within 200 m, then you should be suspicious.

Other sensors and repeat surveys

Ground-based radar provides continuous, high-sensitivity measurements over a small area, and we've previously covered a recent 3D upgrade to ground-based SAR at a mine in PNG. Repeat drone surveys or topographic surveys can also be used, though you have to be honest about the limitations (on most sites they'll only resolve centimetre-level change rather than 10mm). They also need independent checkpoints to be credible, and that's the same principle as why an RTK drone survey still needs checkpoints.

Can InSAR measure 30mm of subsidence? A decision workflow

Yes, it can. The real question is whether this 30mm is credible. Here's a straightforward way of sorting it.

Question If the answer is weak
Is the reference point stable and properly documented? Treat the whole map as relative and unconfirmed
Does the signal continue over many dates rather than just one pair? Suspect atmosphere or an unwrapping slip
Do the ascending and descending passes agree? Suspect horizontal motion or a processing artefact
Is the site vegetated, farmed or seasonally wet? Expect a possible moisture-related bias
Is 30mm clearly greater than the stated uncertainty? Ask for the uncertainty, or do not rely on it
Does any independent measurement support it? Plan a GNSS, levelling or field check

Then sort the result into one of three buckets:

  • Credible: consistent across the various passes and dates, well above the error budget and independently checked.
  • Plausible rather than confirmed (appears physical but no independent confirmation at this stage). Plan a check prior to any decision.
  • Unresolved: fails on more than one of the above questions. Don't act on it, don't dismiss it either.

The third bucket is allowed. To say "this is unresolved, here's what would settle it" is the proper thing to do.

What to take away

A deformation map is a measurement, and so it has an uncertainty and an independent check attached to it just like any other survey deliverable. A line of sight value is not a vertical settlement until you've justified that step. One unwrapping slip can look absolutely identical to the movement you're concerned about.

So when the next map comes in, start with the reference point, the product type, and then what else could possibly generate that number. If you're not sure what you're looking for on the ground or you just want an independent check on the satellite results, our remote sensing team are more than happy to go through it with you. The satellite tells you where to look. The ground has to confirm it.

Back to blog

Leave a comment