Summary: Satellite InSAR and a permanently mounted tilt sensor are usually compared on precision, and the comparison is misleading in both directions. Converted to the same unit they land in the same decade. A tilt is a difference of two displacements divided by the distance between them, so the published 0.7 millimetre per year velocity precision of the Copernicus ground motion service, divided by the 20 metre cell spacing of the Sentinel-1 interferometric wide swath mode, is 0.002 deg, and divided by the 5 metre spacing it is 0.008 deg. Against a tiltmeter’s plus or minus 0.005 deg accuracy figure those are 0.4 and 1.6 times. Precision is therefore not what separates the two methods. Four other things do: the position of the observation point is known to metres while its displacement is known to millimetres, a polar orbit cannot retrieve north-south motion at all, a rate needs a stack of images measured in years, and the free product defines zero as the most stable ground point in the scene. Used as a screening layer over a large area and paired with a borehole array that can see depth, InSAR does something no ground programme can fund. Used as a substitute for a named measurement point, it cannot.

1. The Comparison People Make, and the Number It Leaves Out

The usual framing is a contrast between a satellite that measures millimetres and a sensor that measures hundredths of a degree, with the implication that the two belong to different families and answer to different standards. That framing survives only until the two are converted into the same unit, and the conversion takes one line. A tilt is the difference between two displacements divided by the distance between the two points. Everything then depends on a number that neither instrument publishes, because neither instrument chooses it.

A satellite hands you a displacement and leaves the distance to you: you decide whether to differentiate across one cell, across two, or across a whole embankment, and the tilt you derive changes by exactly that factor. A tilt sensor hands you the angle directly, and its two effective reference points are however far apart its mounting footprint places them, which is an installation decision rather than a specification. So a tilt sensor publishes an accuracy and a resolution, and so does a satellite, and in both cases the figure is a statement about a distance that the customer is silently choosing.

This is not the carrier problem of an earlier article in this series, and it is not the ladder problem of another. In the vibrating wire case the instrument published a resolution in hertz and the conversion factor moved with the reading, so the factor itself had to be computed at the point of use. In the storage tank case every published criterion was already a slope and nothing needed converting at all. Here the two instruments genuinely report in different units, the conversion is a single division, and the whole difficulty is that the divisor is invisible. The article is about what that division produces, and about the four limits that remain after it has been done.

Three instruments are in play rather than two, and keeping them separate is most of the work. The satellite samples a surface over an area and reports a rate. A permanently mounted tilt sensor samples a surface at one point and reports an angle as often as it is told to. An in-place inclinometer array samples the inside of one borehole along its whole depth and reports a profile. Only the third can see a deep-seated failure surface, and that fact decides more monitoring plans than any specification figure in this article.

2. What Sentinel-1 Actually Delivers, in Its Own Numbers

Before any conversion, the published geometry of the mission has to be on the table, because three of the four limits that follow come straight out of it. The figures below are the mission’s own, and the last column is the consequence for a monitoring plan rather than a restatement of the value.

The revisit figure deserves a note because it is a programme fact rather than a specification that a ten-year monitoring plan can be built on. Sentinel-1A launched in April 2014 and Sentinel-1B in April 2016, and the pair provided six-day repeat observations. Sentinel-1B suffered a power supply anomaly in December 2021, the mission ended in August 2022, and the interval returned to twelve days. Sentinel-1C launched in December 2024 and restored a two-satellite constellation, and Sentinel-1D launched in November 2025 and was scheduled to replace Sentinel-1A during June 2026. A monitoring plan that assumes six days should therefore state which constellation it is assuming, and should be able to describe what changes at twelve.

3. The Division Nobody Publishes: Millimetres Into Degrees

Now the conversion. Each row below takes a published precision figure and divides it by the distance over which two of its own measurement points sit, which is the only defensible way to turn a displacement into a tilt. The distances are the mission’s cell spacings, and the tilt sensor rows are included as the reference scale rather than as a conversion. The published velocity figure is itself defined as a difference, which is what makes the division legitimate rather than contrived: the European Ground Motion Service specification states its mean velocity standard deviation of 0.7 millimetres per year at one standard deviation for any pair of points up to 10 kilometres apart, with a constant velocity over the processed interval and a coherence above 0.7. The two points are already in the definition. All that is missing is how far apart the customer wants to place them.

Read the third column against the last two rows. The satellite’s velocity product lands at 0.4 and 1.6 times a tiltmeter’s accuracy figure depending on which cell dimension is used, which is the same decade rather than a different family. Its single-epoch and ground-truth figures land between 2.9 and 16 times that accuracy figure, which is one order of magnitude worse but still the same conversation. Nothing here supports the claim that a satellite is a coarse instrument suitable only for regional subsidence while a tilt sensor is a precision instrument suitable only for a named structure. The claim that does survive is narrower and more useful: the satellite is precise about a rate over a cell, and the tiltmeter is precise about an angle at a mounting footprint.

The factor of four between the first two rows is worth its own sentence, because it is a genuine anisotropy rather than a rounding artefact. The east-west cell is 5 metres and the north-south cell is 20, so a single displacement field differentiated in the two directions gives tilt figures that differ by exactly that factor, and neither is wrong. A monitoring specification that asks for a tilt resolution in degrees therefore has to state the direction it was differentiated in, and most of them do not.

The comparison also produces a result that runs against the obvious conclusion, and it is the reason this article exists. A tiltmeter’s published zero point temperature drift is 0.001 deg per degree Celsius as a typical value, with a maximum of 0.065 deg on the same table. Over an annual swing of 40 degrees Celsius the typical figure accumulates to 0.040 deg. That is 20 times the satellite’s velocity-derived tilt at the 20 metre spacing and 5 times its figure at the 5 metre spacing. So on a slow, deep-seated mechanism, where the satellite is at its best and the ground instrument at its most exposed, the term that decides whether a change can be seen is not the satellite’s resolution and not the tiltmeter’s accuracy. It is the tiltmeter’s zero stability, and it is the last line of the datasheet rather than the first.

4. The Point Is Not the Point

The displacement is known to millimetres, and the point whose displacement it is known to metres. For Sentinel-1 data processed with one proprietary scatterer algorithm, the published case-study position accuracy is plus or minus 8 metres in elevation, plus or minus 8 metres north-south and plus or minus 12 metres east-west. The same figures for a commercial X-band mission are plus or minus 1.5 metres vertically, plus or minus 1 metre north-south and plus or minus 3 metres east-west. Nothing else in this article undermines the attribution of a measurement to a structure as much as that first sentence does.

Put the numbers against real objects. A dam crest ten metres wide occupies half of one north-south cell. A five metre wide wall occupies one east-west cell exactly. And the east-west position error of 12 metres is larger than the plan dimension of many of the structures a monitoring specification names, which means the displacement can be attributed to a neighbouring structure or to the ground between two of them. Dividing that position error by the velocity precision gives the honest shape of the problem: 1 millimetre of displacement carried inside a 12 metre positional uncertainty is a slope of 0.005 deg. That happens to be the same size as the tiltmeter’s accuracy specification, which is a coincidence rather than a relationship, but it is a useful one to hold on to, because it says that the satellite’s imprecision about where the point is sits at the same scale as the finest angle the ground instrument is specified to measure.

There is a second trade hiding behind the product structure, and it is the one that decides which file to order. The products that leave the measurement as a line-of-sight rate keep the full 5 by 20 metre geometry. The products that separate vertical from east-west motion require both an ascending and a descending pass and are then resampled to a 100 metre grid. So the fine geometry and the separated components arrive in different files, and the resolved-component file has lost a factor of twenty in one direction and five in the other before it is used. A specification that asks for vertical movement at five metre resolution is asking for something the operational product does not deliver as one file.

Finally, the floor. A displacement smaller than the cell cannot be resolved at all, because the phase change it produces is averaged with whatever else shares the cell. A crack, a joint, a bolt, a bearing, a single settlement gauge: none of these occupy a cell, and none of them will appear in a satellite product no matter how long the stack or how precise the processing. A tilt sensor’s spatial selectivity is set by the distance between its mounting holes, which on a small bracket is a fraction of a metre and is chosen by the installer. As an illustration of the arithmetic rather than as a product figure, if the two reference points sit 0.2 metres apart then one millimetre of differential movement between them is 0.29 deg, which is 57 times the sensor’s own accuracy figure. The ground instrument is not buying precision at that scale. It is buying the ability to say which part of the structure moved.

5. What the Geometry Cannot See

A radar satellite flies a near-polar orbit and looks sideways rather than straight down, so its line of sight lies in the plane that contains the vertical and the east-west direction. A displacement that is purely north-south is perpendicular to that line of sight and projects onto it as nothing. The table converts the incidence angle range into directional sensitivity, taking the sensitivity to vertical motion as the cosine of the incidence angle and the sensitivity to east-west motion as the sine, with the north-south component set to zero by the geometry rather than by a processing choice.

Two consequences follow, and the second is the important one. The first is a within-swath inconsistency: the same vertical motion reads 26 percent larger in near range than in far range, and the same east-west motion reads 48 percent larger in far range than in near range, so two points at opposite edges of one image are not on the same scale. The second is that whether a site can be monitored by satellite depends on which way its hazard faces. A slope that fails to the east is measurable; the same slope failing to the north is invisible at every incidence angle, because the displacement has no component along the line of sight. The usefulness of the method is therefore a function of the site’s azimuth relative to the orbital plane, and not of the site’s hazard, its rate, or its importance.

There are three honest responses and no fourth. Use two orbits, which separates vertical from east-west but leaves north-south unresolved. Place the radar where the line of sight lies along the expected movement, which is what a ground-based radar unit does by being positioned downhill and in front of the target slope, and which is why ground-based systems trade away coverage to buy the direction that matters. Or accept that on a north-facing slope the satellite cannot serve as the primary sensor and place a permanent ground instrument instead.

6. The Zero Is the Most Stable Point on the Ground

Interferometric radar measures the change in distance between the satellite and the surface, and it measures it differentially, so every result is a difference from a reference. In the free continental product the measurement points are published as line-of-sight velocities referred to a local reference point, and the producer’s own description states how that point is chosen: it is typically the most stable point on the ground as seen from the satellite. It also states the consequence plainly, which is that since nobody knows what that point represents in the physical world, a second product level exists to place the velocities against something with a physical meaning.

Read that once more, because it is the least intuitive line in the whole subject. The map is defined as zero at the place that moves least, chosen by the measurement itself. If an entire district subsides together, the most stable point is still whatever point moved least, and the district’s absolute movement is not what the basic product reports. The higher product level addresses exactly this: it takes the same velocities and ties them to a model derived from global navigation satellite system data, built from 4,716 permanent stations across Europe on a 50 kilometre grid, with typical model standard deviations of 0.1 to 0.2 millimetres per year horizontally and 0.5 millimetres per year vertically. The result is an absolute velocity in a known reference frame rather than a relative one.

Two further relativities are worth knowing before a specification is written. The temporal reference differs between datasets, so a displacement series is a displacement since the reference time of that particular delivery rather than since commissioning. And the published epochs span about five years with an annual update, so the free product answers the question how fast on average over five years, and does not answer the question what happened last month.

This is the satellite instance of a problem this series has already met twice, once for dams and once for storage tanks, and the treatment is the same here as there, which is why it is stated rather than re-derived. A relative product must never be allowed to define zero. Choose a reference on a physical argument, say out loud which product level the velocities came from, and where a national service offers a calibrated level, use it and keep the benchmark you can put a hand on underneath it. The free product is a screening layer. A benchmark is a benchmark.

7. A Rate Needs Years, and the Failure Does Not Wait

The precision figures in the third section are velocity precisions, and a velocity is a fit through a stack. That means the headline number is not available on the day the satellite passes; it is available after enough images have accumulated for the fit to be worth making. The published guidance puts a usable scatterer solution at more than ten acquisitions, a precision of about half a millimetre per year at roughly thirty, and the widely quoted figure of about one millimetre per year at more than a hundred images over three years.

The last row is the one to carry into a specification review, because it shows why a satellite accuracy figure cannot be quoted without the movement it was measured against. On a site creeping at 50 millimetres per year, a 7 millimetre agreement is good news. On a site creeping at 5 millimetres per year the same agreement is larger than the annual signal, which means the result cannot distinguish one year of movement from one year of noise. The peer-reviewed study that produced that figure drew its conclusions carefully and they are worth reproducing, because they come from a group that had a ground-truth site and five failure cases to compare against. Their summary is that the technique can serve as a valuable hazard screening tool across large, multiple, inactive or legacy tailings facilities, that most potential failure modes may not exhibit accelerations in precursor deformation that the method can detect, that the revisit interval prevents detection of instantaneous failure mechanisms, and that long-term monitoring programmes for tailings dams should be integrated with a combination of remote sensing and field instrumentation.

Two further sentences in the same literature deserve to be passed on rather than paraphrased away. The first is that the benefit of hindsight is an important factor in how pre-failure data is perceived in forensic studies, which is a polite way of saying that a signal which looks clear after a collapse is much harder to identify before one. The second is a paper title rather than a result: a study on the perspectives on the prediction of catastrophic slope failures from satellite data, cited in the same literature, is by authors who also supplied the published limits of failure time prediction from a ground instrument. That symmetry is the point. The researchers who told the industry what inverse velocity cannot do have also told it what a satellite cannot do, and neither paper argues for abandoning the other method.

8. The Depth Nobody Measures From Above

Each of the three methods samples a different thing, and the differences are not matters of quality. The table is the whole argument for using more than one of them, and the last column is the one that decides most specifications.

The reason an array rather than a surface instrument is standard on a tailings storage facility is not cost and not habit. An array is permanently installed in a borehole and reads the profile at fixed depth intervals, which is what makes slow deep-seated deformation detectable at all. A deep-seated failure surface can move a considerable distance at depth while the surface above it barely rotates, and an angle measured on the surface cannot distinguish that from nothing happening. It is a limit of the measurement, not of the instrument, and no amount of resolution in degrees repairs it.

That is the shape of the arrangement at the Martabe Gold Mine in North Sumatra, where Agincourt Resources operates the Purnama open pit and its tailings storage facility. The facility was designed to International Commission on Large Dams standards and approved by the Dam Safety Commission, and continuous monitoring is essential regardless, because embankments deform over time through settlement, seasonal water level changes and incremental loading as new tailings are deposited. The system installed in 2020 is ZC Sensor’s wireless in-place inclinometer array, which reads the embankment profile and uploads to the mine’s cloud data centre without field visits, feeding the operator’s zero-accident safety programme.

Put the coverage figures next to the depth figure and the trade becomes arithmetic rather than opinion. On the company’s own published slope deployment at an iron mine dump, the array was four boreholes drilled 40 metres deep with eight sensing units in each at 5 metre spacing, which is 32 sensing points distributed along four vertical lines. The satellite, at one measurement point per 5 by 20 metre cell, could carry up to about 10,000 points across a single square kilometre, and a recent study area on the Florida coastline covered 1,838 square kilometres. So the satellite buys coverage with the sampled depth set to zero, and the array buys depth with the coverage set to four lines. A monitoring plan that contains only one of them has chosen which question it will not answer.

9. A Decision Order That Uses Both

The methods are not competing for the same job, and the practical problem is sequence rather than selection. The four questions below are the ones that decide it, in the order they should be asked.

Five rules follow, and each one is a sentence that can be put into a monitoring specification without further explanation. First, never compare a millimetre with a degree until both have been divided by the distance they were measured over, and write that distance down, because the satellite does not choose it, the sensor does not choose it, and the reading changes by the factor between one choice and another. Second, read the position accuracy and not only the displacement precision, because a millimetre at a point known to 12 metres is not a measurement of the point you named. Third, check the site’s aspect against the orbital plane before buying a satellite product, because a north-facing slope is invisible to a polar orbit at every incidence angle. Fourth, treat the default reference of a free product as a statistical convention rather than a datum, and keep a benchmark underneath it. Fifth, state the time base, and be prepared to explain why a rate fitted over five years cannot be updated into a warning this month.

One thing should be said plainly about where this article comes from. ZC Sensor publishes ground instruments: wireless tiltmeters, in-place inclinometer arrays and their loggers, and the deployments cited here are the company’s own. The company does not sell satellite data, and every figure attributed to Sentinel-1, to the Copernicus ground motion service or to a published case study in this article comes from the mission documentation, the service specification or the peer-reviewed literature that is named alongside it. The conclusion the two sides of this article actually support is not that one method beats the other. It is that the satellite is the only one of the three that can be pointed at a whole portfolio at once, and that the instrument which can tell you at what depth the ground is moving is the one you have to drill for.

Frequently Asked Questions

1. Can satellite InSAR replace tilt sensors?For screening a large area it does something no ground instrument can, and for confirming what is happening at a named point it cannot. Converted to the same unit, the satellite’s velocity precision is 0.4 to 1.6 times a tilt sensor’s accuracy figure depending on the cell dimension you differentiate over, so precision is not the difference. The differences are that the satellite measures a cell 5 by 20 metres rather than a mounting footprint, that it cannot retrieve north-south motion at all, that its rate precision needs years of images, and that its default zero is the most stable ground point in the scene. The published purpose of the method in the tailings literature is hazard screening rather than surveillance.
2. Why can InSAR not measure north-south movement?Because the satellite flies a near-polar orbit and looks sideways rather than straight down, so its line of sight lies in the vertical and east-west plane. A displacement that is purely north-south is perpendicular to that line of sight and projects onto it as zero, at every incidence angle between 29.1 and 46.0 degrees. Vertical motion is still measured, at 0.874 of its true size in near range and 0.695 in far range, and combining an ascending pass with a descending one separates vertical from east-west, but no combination of orbits recovers the north-south component. The practical consequence is that whether a slope can be monitored from orbit depends on which way it faces rather than on how dangerous it is.
3. How do I compare an InSAR millimetre with my tilt sensor’s degree?Divide. A tilt is a difference of two displacements divided by the distance between them, so a published velocity precision of 0.7 millimetres per year divided by the 20 metre north-south cell spacing is 0.002 deg, and divided by the 5 metre east-west spacing it is 0.008 deg. Against an accuracy figure of plus or minus 0.005 deg those are 0.4 and 1.6 times, which is the same decade rather than a different order of magnitude. The step that matters is stating which distance was used, because it changes the answer by a factor of four within a single image, and because neither the satellite nor the sensor chooses it.
4. Do I need both instruments?Only if the mechanism operates at more than one depth or on more than one time base, which on a tailings facility, a large embankment or a subsiding district is almost always the case. A deep-seated failure surface moves the surface hardly at all, so no surface method of either kind settles it, and the measurement that does is an array inside a borehole reading the profile at fixed depth intervals. What the satellite adds is coverage: one measurement point per 5 by 20 metre cell across a whole site, and across every site an operator owns, which no borehole programme can fund. The usual arrangement is screening from above and confirmation from inside, with the ground instrument also providing the zero that the free product cannot.