InSAR Explained: What Satellite Radar Can and Cannot Measure on the Ground

Srihari Maddula

8 hours ago

16 min read

Earth seen from space with a satellite — Photo: NASA / Unsplash

Satellite radar can see the ground move by a few millimetres from 700 kilometres up. That sentence is true, and it is also the source of most of the disappointment people feel after their first InSAR project, because it leaves out three conditions: where, how fast and in which direction.

InSAR stands for interferometric synthetic aperture radar. It compares the phase of radar echoes from the same patch of ground on two or more dates, and turns the difference into a measure of how far the surface moved toward or away from the satellite. Used well, it gives wide-area, repeat, low-cost monitoring of slow ground movement: subsidence of cities and reclaimed land, settling around reservoirs and mines, creep on slopes, uplift near injection wells. Used carelessly, it produces beautiful, confident maps of noise.

This post is a myth-buster in eight rounds. Each takes a claim you will hear from vendors, journalists and colleagues, says what is true in it, shows where it breaks, and gives a test you can run. The arithmetic is done in code, so you can check every number. Where I rely on a public source, I say which kind, and where something changed in 2026, I flag it, because the Sentinel-1 constellation was reorganised this year and many guides online are out of date.

What InSAR Actually Measures

A radar satellite emits a pulse and records the echo, including its phase, which is a number between zero and a full cycle describing where in a wave the echo arrived. If the ground has moved toward the satellite between two passes, the path is shorter and the phase shifts. One full cycle of phase corresponds to half a wavelength of movement, because the signal travels out and back.

For the Sentinel-1 satellites, which work in the C band at a wavelength of about 5.66 centimetres, one full cycle of phase, called a fringe, is 2.83 centimetres of movement along the line of sight. Phase can be measured to a small fraction of a cycle, which is how the millimetre claim arises.

import math
WL = 5.66                                    # Sentinel-1 C-band wavelength, centimetres

def los_from_phase(phi_rad, wl_cm=WL):
    return phi_rad * wl_cm / (4 * math.pi)

print(round(los_from_phase(2 * math.pi), 2))     # 2.83 cm per fringe
print(round(los_from_phase(0.1) * 10, 2))        # 0.45 mm for 0.1 radian of phase

Three words in the definition do the work: toward or away, along the line of sight, and between two dates. InSAR measures relative movement of a patch of ground against a reference, in one direction, between acquisitions. Every myth below is a way of forgetting one of those.

Myth 1: Satellites Measure Millimetres Everywhere

What is true. On a good target, such as a building, a bridge, a road or bare rock, a long series of images can detect a few millimetres a year of movement.

Where it breaks. The measurement requires that the radar echo from each pixel be stable from one date to the next, a property called coherence. Where the surface changes between passes, the phase becomes random and the pixel is useless. Farmland that is ploughed, irrigated or harvested, dense forest, water and wet soil lose coherence quickly, and in much of India those are the landscapes people care about.

The noise in a single interferogram depends on coherence and on how many neighbouring pixels are averaged, called looks. The code below uses the standard approximation.

def phase_std(coh, looks):
    return math.sqrt(1 - coh ** 2) / (coh * math.sqrt(2 * looks))

for coh in (0.8, 0.5, 0.3):
    s = phase_std(coh, 50)
    print(coh, round(s, 3), round(los_from_phase(s) * 10, 2))
# 0.8 -> 0.075 rad, 0.34 mm   0.5 -> 0.173 rad, 0.78 mm   0.3 -> 0.318 rad, 1.43 mm

Even at low coherence, the nominal noise looks small, around a millimetre or two. That is the trap. The formula describes random noise on pixels that still carry a signal. Below some coherence the signal is lost entirely, unwrapping fails and the pixel is masked, so the map has holes where the farmland is. Agencies that publish displacement products flag low-coherence areas as unreliable for exactly this reason, and exclude them from their recommended mask.

The test. Ask for the coherence map beside the displacement map, and ask what share of your area of interest is coherent on most dates. If the answer is 20 percent, you have a 20 percent map.

A displacement map is a map of where radar could see, with movement painted on it. Look at the coverage before you look at the colours.

Myth 2: InSAR Measures Vertical Subsidence

What is true. If the ground sinks, the satellite sees it, and subsidence is the most common application.

Where it breaks. The satellite measures along its line of sight, which is tilted from the vertical by the incidence angle, roughly 30 to 46 degrees for Sentinel-1. A purely vertical movement is seen only as the vertical movement multiplied by the cosine of that angle. A movement toward the east or west is also partly visible, and a movement along the satellite's track is invisible.

def vertical_to_los(v, inc_deg):
    return v * math.cos(math.radians(inc_deg))

print(round(vertical_to_los(10, 39), 2))       # 7.77 mm along the line of sight for 10 mm of sinking

A map in line-of-sight units is not a map of subsidence. To convert it, you assume the movement is vertical, which is often reasonable for subsidence and wrong for a slope. To separate vertical from east-west movement, you combine ascending and descending passes, which look at the ground from opposite sides, and solve for the two components.

The test. Ask what units the deliverable is in and what assumption was made to convert it. If the answer is that the vendor simply divided by the cosine of the incidence angle, ask whether the site is a subsiding plain or a moving slope.

Myth 3: A Six-Day Revisit Means We Will See Everything That Moves

What is true. Sentinel-1 revisits the same ground about every six days with two satellites. Until June 2026 that pair was Sentinel-1A and 1C. Sentinel-1D, launched in November 2025, opened its data to users on 17 April 2026, and Sentinel-1A operations were terminated on 29 June 2026, leaving 1C and 1D as the long-term constellation. The orbital shift moved the six-day pattern by a day relative to the old A and B timeline, a detail that matters if you compare series across the change.

Where it breaks. Frequent revisits do not help if the ground moves too fast between two of them. The phase repeats every half wavelength, so the difference in movement between neighbouring pixels must stay under a quarter wavelength per interval, or the fringes cannot be counted and the map is ambiguous.

print(round(WL / 4, 3))                               # 1.415 cm between adjacent pixels per interval
print(round(WL / 4 * 10 / 6, 2))                      # 2.36 mm per day at a 6-day revisit

That limit, about 1.4 centimetres of differential movement between adjacent pixels in six days, rules out some of the situations people most want to monitor: sudden collapse above shallow mine workings, a fast landslide, a sinkhole opening. Slow creep before such an event may be seen. The event itself usually is not.

The test. Estimate the rate and the spatial gradient you expect, not only the total movement. If the answer exceeds a centimetre or so between neighbouring pixels per revisit, InSAR with C-band is the wrong tool for the fast phase, and you need other instruments.

Myth 4: Radar Sees Through Clouds, So the Atmosphere Is Not a Problem

What is true. Radar penetrates cloud and works at night, which is why it suits the monsoon better than optical imaging.

Where it breaks. The atmosphere still delays the signal. Variations in water vapour between two dates add phase that has nothing to do with ground movement and can look like several centimetres of apparent displacement across a scene, with smooth patterns that mimic real deformation. The tropics and the monsoon season are among the worst conditions.

Time-series methods reduce the problem. The atmospheric signal is correlated in space and random in time, so averaging many dates and filtering the spatially smooth component separates it from steady movement. Weather-model corrections help further. Stacking a year of acquisitions gives better precision than any single pair.

The test. Ask how atmospheric effects were handled, and ask for an estimate of the uncertainty of the velocity in millimetres per year, not only the velocity. A vendor that cannot give an uncertainty has not estimated one.

Myth 5: More Images Always Mean More Accuracy

What is true. A long time series improves precision, because random errors average out and steady movement accumulates.

Where it breaks. Coherence decays as the time between two images grows, because the surface changes. The decay is faster over vegetation and wet soil than over built-up land, which is why short intervals are valuable and long ones may contain nothing usable.

def coherence_decay(t_days, tau_days, floor=0.0):
    return floor + (1 - floor) * math.exp(-t_days / tau_days)

for tau in (60, 15):                     # illustrative decay constants for stable and changeable ground
    print(tau, [round(coherence_decay(t, tau), 2) for t in (6, 12, 24, 48)])
# 60 -> [0.9, 0.82, 0.67, 0.45]    15 -> [0.67, 0.45, 0.2, 0.04]

With a decay constant of 60 days, pairs at 24 days still hold a coherence of 0.67. With 15 days, the 24-day pair has fallen to 0.2 and is of little use. The values are illustrative, and the real constant has to be measured for your ground and season. Time-series methods choose the pairs with this in mind, and they deal with a subtler problem too: small phase biases from soil moisture and vegetation that accumulate in long series and can look like slow movement.

The test. Ask which pairs were used and how phase bias was handled. A series built from every possible pair without regard to coherence is not more accurate, only bigger.

Myth 6: Free Data Means a Cheap Answer

What is true. The raw radar data from the Sentinel missions is free and open, and agencies and data centres now offer processed displacement products over some regions.

Where it breaks. The data is the cheap part. Turning it into a trustworthy displacement time series requires processing expertise, choices about pairs, filtering, atmospheric correction, reference points and unwrapping, and then validation against something independent. Published displacement products exist for some regions, but coverage is geographically limited. The agency running one such product tells users that building a similar workflow for other regions is a real task.

A reasonable budget for an InSAR monitoring project is mostly skilled time: setting up, processing, checking, interpreting and reporting. The satellite time is free, and the answer is not.

The test. Ask what is in the price. If the quote is a small fee for a map, ask whether it includes validation and an uncertainty estimate. If it is a large fee, ask what you are buying that a processed open product would not give you.

Myth 7: InSAR Replaces Ground Instruments

What is true. InSAR covers areas that no ground network could, with a repeat schedule and no site access, at a low cost per square kilometre.

Where it breaks. InSAR gives relative movement, along the line of sight, at the dates of the passes. A ground receiver using satellite navigation gives absolute three-dimensional position continuously. A precise levelling line gives millimetre-scale heights on a schedule you choose. A tiltmeter, an extensometer or a piezometer gives readings in hours or minutes, in the places that matter.

The strongest projects combine them. InSAR finds where to look across a whole district. A small number of ground instruments calibrate it, tie it to an absolute reference and fill the gaps in time. Trigger levels for action are set from the fast instruments, and the satellite map shows the pattern.

The test. Ask what the monitoring is for. If a decision must be made within hours, the satellite alone is too slow. If the question is which of fifty sites is moving, the satellite is the cheapest way to start.

Myth 8: The New L-Band Satellite Will Fix Everything

What is true. NISAR, the joint NASA and ISRO mission launched in 2025, carries an L-band radar with a wavelength of about 24 centimetres and a secondary S-band radar. Longer wavelengths penetrate vegetation better and lose coherence more slowly, so L-band should see ground where C-band cannot. A comparison presented at a 2026 European Space Agency workshop found higher correlation over vegetated regions with NISAR, and found that Sentinel-1 missed deformation in a non-vegetated area where it decorrelated.

Where it breaks. The same comparison reported that many NISAR scenes suffer from ionospheric artefacts, which are not correctable to date, while Sentinel-1 largely does not. NISAR's revisit is about 12 days, twice the Sentinel-1 interval, which limits the fast movements it can follow. And the Sentinel-1 archive has a long head start, with a time series going back over a decade.

The realistic position is complementary. C-band gives frequent, long, well-corrected series where the ground is coherent, and L-band extends coverage into vegetated land. Projects that can use both will see more of the ground than either alone, and the processing of the second is still maturing.

The test. Ask whether the vendor can use both wavelengths and what each is expected to add for your site.

Ascending, Descending and the Geometry of a Site

Sentinel-1 looks to the right of its track. On its northbound, ascending passes it looks roughly eastward, and on its southbound, descending passes it looks roughly westward. The same patch of ground is therefore seen from two directions, with two different line-of-sight vectors.

That matters in three ways. First, a movement is visible in each geometry according to its alignment with the line of sight. Ground moving east is seen as movement toward the satellite on one pass and away from it on the other, so comparing the two separates horizontal from vertical movement. Second, steep terrain distorts radar images. Slopes that face the satellite are foreshortened or laid over onto the terrain in front of them, and slopes that face away are in shadow, with no echo at all. A hillside that is invisible from one pass may be visible from the other. Third, the incidence angle sets the sensitivity: a steeper, more vertical look is more sensitive to vertical movement, and a shallower look more sensitive to horizontal.

Before you commission a project, ask for a geometry check of your site: which passes see it, at what incidence angle, with what layover and shadow. A site on the wrong side of a ridge can be a site that one geometry cannot monitor at all.

Always ask which geometry sees your site, and what it cannot see from there. A shadowed slope is not stable, only unobserved.

Two Ways to Process: Persistent Scatterers and Small Baselines

Two families of processing methods turn a stack of radar images into a displacement time series, and they suit different ground.

Persistent scatterer methods look for individual pixels that stay bright and stable on every date, such as corners of buildings, metal structures, bridges and exposed rock. They use the full stack, including long time gaps, because a stable point remains coherent however much time passes. They give very precise measurements at those points and none between them. They suit cities and infrastructure, where such points are plentiful.

Small baseline methods work on distributed ground, such as open land, scrub and thin vegetation, where no single pixel is a stable target. They form many interferograms from images close in time and in orbit position, average neighbouring pixels to reduce noise and combine the results into a time series. They cover areas that persistent scatterer methods leave blank, at lower precision per pixel.

Method family	Best ground	Strength	Weakness
Persistent scatterers	Cities, infrastructure, rock	High precision at stable points, works across long gaps	No data between points, needs many images
Small baselines	Open land, mixed cover	Wider coverage, uses short pairs where coherence holds	Phase bias in long series, lower precision
Combined approaches	Mixed sites	Uses both kinds of target	More processing, more choices to defend

Many recent workflows combine the two, estimating phase over groups of pixels and then correcting the bias that soil moisture and vegetation introduce. A provider should be able to say which approach it used for your site and why.

What InSAR Does Well

Set the myths aside and a clear picture remains. InSAR is excellent for slow, wide-area movement on stable, coherent ground, measured against a well-chosen reference.

Application	Why it suits	Main caution
Urban subsidence from groundwater withdrawal	Buildings and roads are coherent, movement is slow	Separating seasonal from long-term movement
Reclaimed land and port areas	Hard surfaces, settling over years	Reference point choice, pier and quay movement
Reservoirs, dams and embankments	Concrete and rock, slow deformation	Vegetation on embankments
Mine subsidence over time	Wide area, slow phase	Fast collapse and vegetated spoil
Slope creep before failure	Early warning of slow movement	Fast failure, line-of-sight geometry on steep slopes
Infrastructure corridors, rail and pipelines	Long linear targets, regular revisit	Narrow targets and low coherence in cuttings

Studies of cities with heavy groundwater pumping, including cities in India, have reported measurable subsidence using this technique. Whether it applies to a particular site is a question for the site, not for the headline.

Choosing a Reference Point

Every InSAR velocity is relative to something. The processing picks a reference, a pixel or a small area assumed not to move, and reports everything else against it. A poor reference shifts the whole map.

A good reference has four properties. It is coherent on every date of the series, so it must not be under vegetation, water or farmland. It is far from the suspected deformation, so it does not share it. It is on stable ground, ideally rock or a deep-founded structure. And it is independently confirmed, with a ground receiver or a levelling benchmark where possible.

Check the reference by moving it. Recompute the map with a second reference on different stable ground. If the pattern holds and only the overall offset changes, the result is robust. If the pattern changes, the reference was part of the signal.

Move the reference point and look at the map again. A result that survives the change is evidence. One that does not was a property of the reference.

How to Read an InSAR Deliverable

A good deliverable is a package, not a picture. Check seven things.

The area covered and the area that is actually coherent, shown as a mask. The reference point, and evidence that it is stable. The units and the direction, line of sight or converted, with the assumption stated. The period and the acquisitions used, including which satellites, since the Sentinel-1 constellation changed in 2026. The velocity with an uncertainty in millimetres per year. The treatment of the atmosphere and of phase bias. And the validation, whether against ground instruments or against an independent product.

A map that shows velocity only, with no mask and no uncertainty, is a picture. A map that shows all seven is evidence.

What to Ask a Provider

Question	Why it matters	A good answer
What share of my site is coherent, and when	Sets the real coverage	A coherence map and a percentage by season
What is the expected uncertainty in millimetres per year	Separates signal from noise	A number with a method
How do you handle atmosphere and phase bias	The commonest hidden error	Named methods, plus validation
Are results in line of sight or converted	Units decide interpretation	Stated, with the assumption
Which satellites and dates	Constellation changed in 2026	A list, with treatment of the change
How will you validate	Independent confirmation	Ground instruments or an independent product
What can it not see here	Honesty about limits	A specific list for your site

A Composite: A Reclaimed Port Terminal

This is a representative composite, genericised, from several port and infrastructure discussions. The numbers are illustrative.

A port operator wanted to know whether a recently reclaimed terminal area was settling evenly. The quay structures were concrete, the yard was paved with container stacking areas, and part of the area was open fill with scrub growth.

The first pass was a quick open-product look. Across the paved and built areas, coherence was high on almost every date, and the time series showed a slow settling of a few millimetres per month in the fill zone, with a clear boundary at the quay wall. Across the scrub, coherence was poor, and the map showed holes.

The operator's engineers asked two questions that shaped the project. First, how much of the movement was the reference point itself? The reference had been chosen on a building near the gate, and a check against two ground benchmarks showed that it was moving slightly, which shifted every velocity by about the same amount. Second, was the movement vertical? The ascending and descending passes disagreed slightly near the quay wall, which indicated a horizontal component toward the water that a vertical-only assumption would have missed.

The final package combined both passes, used a reference tied to the benchmarks, flagged the scrub as unmonitored, and added four ground instruments in the fill zone to give daily readings and trigger levels. The satellite map told the operator where to look. The ground instruments told it when to act.

A Ninety-Day Plan for a First InSAR Project

A first project is easier with a fixed order of work. This sequence fits a typical site and a small team.

Days	Task	Output
1 to 10	Define the question, the area, the expected rate and the decision the result will inform	A one-page brief
10 to 20	Check geometry, available acquisitions and expected coherence for each pass	A feasibility note with a coverage estimate
20 to 40	Process a first stack, using a published workflow, with an explicit reference choice	Velocity map, coherence mask and uncertainty estimate
40 to 55	Compare with any ground data, including benchmarks, receivers or a published product	A validation table
55 to 70	Repeat with an alternative reference and an alternative processing choice	A sensitivity check
70 to 85	Interpret with an engineer or geologist who knows the site	A written interpretation, with what the data cannot say
85 to 90	Decide on ground instruments, repeat schedule and triggers	A monitoring plan

Each step has a decision attached. If the feasibility note shows coverage under a third of the area, rethink the project before spending time on processing. If the validation shows disagreement with ground data, find out why before drawing any conclusions. If the interpretation depends on a single pixel or a single date, treat it as a hypothesis.

A first project should end with a decision about monitoring, not only a map. The map is a means to that decision.

What a First Project Takes: Time, Skills and Budget

A first InSAR project is mostly skilled time, and it is useful to see where the hours go before you ask for a quote or plan a team.

Setting up takes the first stretch: defining the question, checking geometry and choosing the data. Processing is the part most people picture, but with modern workflows and open tools it is often the shortest step. Checking is long: comparing against ground data, testing the reference, examining masks and looking for artefacts. Interpretation is the step that needs someone who knows the site, whether a geotechnical engineer, a mining engineer or a hydrogeologist. Reporting, with the seven items listed earlier, completes the work.

Activity	Share of effort	Skill needed
Framing and feasibility	About a tenth	Remote sensing and domain knowledge
Data selection and processing	About a quarter	Radar processing
Checking and validation	About a third	Radar processing plus ground survey knowledge
Interpretation	About a quarter	Domain engineering
Reporting and handover	About a tenth	Both

The shares are indicative, taken from the way such projects typically divide, and they vary with site and scope. The pattern is more useful than the figures: the largest costs are checking and interpretation, which are the steps a cheap quotation tends to omit.

Decide early whether to buy a service or build a capability. A single site on a single question is usually cheaper as a service. A programme of many sites over years, such as a port authority monitoring its estates or a mining company watching several leases, justifies a team that owns the workflow, using open data and published methods, and uses a specialist for peer review.

Seven Questions Before You Commission a Project

What is the decision this monitoring will inform, and how fast must it be made?
How fast and how steeply do I expect the ground to move?
What is the surface: built, rocky, farmed, forested or wet?
Which direction will the movement be in, and can one geometry see it?
What ground truth exists, or can be installed?
What uncertainty would change my decision?
Who will interpret the result, and are they independent of the seller?

The Short Version of Eight Myths

InSAR is a precise instrument with a narrow envelope. It measures relative movement along one direction, on coherent ground, slower than a centimetre between neighbouring pixels per revisit, against a reference you choose. Inside that envelope it is remarkable value. Outside it, it produces confident noise.

Ask for the coverage, the units, the uncertainty and the validation. Combine the satellite with a few ground instruments where a decision depends on it. And remember that the constellation, the satellites and the processing products all moved in 2026.

Related service: Geospatial Intelligence Solutions

EurthTech is an engineering-first product design and embedded systems company with a development center in Hyderabad. To discuss a project, write to connect@eurthtech.com or see https://www.eurthtech.com/case-studies.