How Radar Sees

A camera needs sunlight and clear air. Radar brings its own light — it sends a radio pulse at the ground and listens for the faint echo that returns. From that echo alone it can build a picture sharp enough to pick out a vehicle, day or night, through cloud and storm. Over the course of this essay we'll build that idea up from first principles, one interactive figure at a time, until we can ask a serious question about the thousands of Starlink satellites already overhead: could they quietly do the same thing?

We start here, in Part 1, with the fundamentals of seeing: the echo, the target, how sharp the picture can be, and how strong. Later parts build the image (Part 2), catch things that move (Part 3), take apart Starlink's own radios (Part 4), and finally put everything together (Part 5). Nothing along the way has to be taken on faith — every number the conclusion rests on is one you'll be able to compute yourself.

Every figure is live — drag them, pull the sliders. Animations run by default; each one has a small pause button in its top-right corner if you'd rather it hold still or save power.

Meet the constellation

Starlink is a low-Earth-orbit mega-constellation: thousands of small satellites circling about 550 km up — low enough that each races over the horizon in a few minutes, so it takes a whole swarm to keep any one spot on Earth covered. Each spacecraft is flat-packed to stack dozens per rocket, unfolds its solar arrays, and points a phased-array antenna down at the ground. They relay your data to pizza-box terminals below and to each other through laser links, handing you off from one satellite to the next as they streak past at 7.6 km/s.

Let's start by looking at the constellation itself. The figure below is not an artist's impression: it downloads today's published orbital elements for every tracked Starlink satellite and computes where each one is at this moment — more than eleven thousand spacecraft as of August 2026, each steering antenna beams at the ground. Drag to spin the globe.

The count is the unprecedented part. In 2019 Starlink was a few dozen satellites; today it is the eleven thousand you just saw overhead — by far the largest constellation ever flown, and still climbing by the thousands each year. SpaceX is licensed for about twelve thousand and has filed for as many as forty-two thousand. Drag the year to watch it grow.

SpaceX, it should be said, already builds spy satellites. A separate classified line called Starshield flies on Starlink-derived buses for the National Reconnaissance Office — a $1.8 billion contract, more than 150 satellites, and signals that radio hobbyists have already fingerprinted from the ground. Whether SpaceX would build a surveillance constellation is therefore not in question — it already does, openly. The question this essay asks is whether the commercial fleet, the one sold as internet access, could quietly serve as one too.

The fleet itself has moved through generations — the early v1.5, today's workhorse V2 Mini, and the giant new V3 that rides Starship to orbit and can talk straight to an ordinary phone. Each generation is larger, with a bigger antenna and more power. Here are all three drawn to the same scale, next to a person — drag to look around them.

Each of these spacecraft was built to sell internet access. But a large, powerful, electronically steered antenna pointed at the ground is also the heart of a radar, and over the rest of this essay we'll work out, from first principles, how far that resemblance goes. We begin the way every radar begins: send a single pulse, and time the echo.

Part 1Seeing with Echoes

One pulse, one echo

Stripped of everything else, radar is a stopwatch. The antenna emits a short radio pulse; the pulse travels outward at the speed of light, strikes something, and a faint echo returns. The radar measures a single quantity — the delay Δt between the pulse and its echo. And because the pulse covered the distance twice, out and back, the target sits at

R = c·Δt / 2

— half of what the light travelled; the 2 accounts for the round trip. The whole idea of radar is contained in that little equation: distance, measured with a clock. Every picture in this essay, however elaborate, is this one measurement made millions of times. Drag the target and watch its echo slide along the time axis.

The idea is ninety years old. In 1935 Robert Watson-Watt bounced the BBC's shortwave transmitter at Daventry off a passing bomber to prove that aircraft could be caught by radio alone; within four years Britain had ringed its coast with the Chain Home towers that helped win the Battle of Britain. Everything since — synthetic apertures, satellites, the whole of this essay — is that same pulse and echo, refined.

The view from orbit

Now let's put the stopwatch in orbit. A satellite races along its orbit while it looks sideways at the ground — never straight down. The angle between the ground and the line up to the satellite is the grazing angle, and the distance the pulse travels to reach a target is the slant range. Drag the scene, and notice how lowering the satellite or flattening the grazing angle changes that slant range — the single most important distance in all of radar.

Why does the slant range matter so much? Because the pulse spreads out as it travels, weakening with the square of the distance on the way out — and, after it bounces, again on the way back. The round-trip loss goes as the fourth power of range. Double the distance and the returning echo is sixteen times fainter.

Let's put numbers on it. Across a 500 km slant range the round trip takes about 3.3 milliseconds — the stopwatch part is easy. The power is not: transmit a one-kilowatt pulse and the echo comes back at around 10⁻¹⁶ watts, nineteen orders of magnitude below what left the antenna. Most of radar engineering is the business of hearing that echo anyway.

What makes a target bright

When the pulse reaches the ground it scatters. How much energy comes back depends on the surface: its material, its roughness, and the angle it presents. Radar people fold this into a number called sigma-zero (σ⁰) — the radar reflectivity per unit area. It is why a radar image can look eerily like a photograph — a grainy one, speckled by the interference of its own echoes — even though no sunlight was involved: the radar brought its own illumination.

Radar arithmetic runs in decibels (dB) — a ratio on a logarithmic scale. 3 dB is roughly a factor of two, 10 dB exactly ten, 20 dB a hundred; negative values are fractions. So when a figure says the ground's σ⁰ is −15 dB, it means each unit of area returns about 3% of what falls on it.

Smooth water is a mirror: it bounces the pulse away from the antenna and looks nearly black. A stealth aircraft is dark on purpose: its facets are shaped to fling the pulse anywhere but back at the sender, and absorbent coatings mop up what the shaping misses. It is a defence built for one geometry — it works because the transmitter and the receiver stand in the same place. But metal with sharp corners blazes, and — surprisingly — so do humans and animals, full of salty, conductive water and shaped in lumpy, echo-friendly ways. And it all depends on wavelength: short X-band waves scatter off leaves, while longer L-band waves sail through a forest canopy to image the ground beneath. Pick a surface and a band, and compare.

How sharp is the picture?

Two targets sitting close together only show up as two things if their echoes arrive at distinguishably different times. Along the radar's line of sight, what separates them is not power and not antenna size — it is the bandwidth of the pulse. The finest distance radar can resolve in that direction is

δ_r = c / 2B

where c is the speed of light and B is the bandwidth. More bandwidth, finer detail — and, remarkably, this does not depend on range at all. A radar a thousand kilometres away resolves detail just as finely as one next door, given the same bandwidth. That may seem to contradict the fourth-power law of the last section, but the two laws govern different things: distance never blurs the picture, it only dims it. What range costs is sensitivity, and sensitivity is the next section's business.

But where does bandwidth come from? Not from shortening the pulse: a 25 cm cell corresponds to a pulse 1.7 nanoseconds long, which carries almost no energy — its echo would drown instantly. So the radar transmits a long pulse and sweeps its frequency as it goes — a chirp — then, on receive, runs the echo through a matched filter that piles the whole sweep up into one sharp spike. The spike behaves exactly as if a crisp nanosecond pulse had been sent, but it carries the energy of the long one. This is pulse compression, and every imaging radar does it.

In the figure, two targets sit a settable distance apart. Widen the bandwidth and watch their blurred returns sharpen and split into two distinct peaks the instant the spacing exceeds the resolution cell.

That last slider hides a subtlety. Resolution along the slant direction is fixed by bandwidth, but what we care about is resolution on the ground, which depends on the grazing angle: the ground cell is δ_r / cos(grazing). Push toward looking straight down and it blows up toward infinity. That is why a radar satellite always looks off to the side — and why the geometry in the first figure was never pointed at nadir.

Everything in this section is one axis of the picture — the across-track axis, distance measured along the beam. The other axis, along the satellite's flight path, is a different problem entirely: no amount of bandwidth helps there, and the antenna's size owns it. Sharpening that axis is the synthetic-aperture trick, and all of Part 2 is devoted to it.

For scale: crisp 25 cm detail in the range direction — enough to identify a military vehicle — takes about 600 MHz of bandwidth. That number will come back when we ask what Starlink's antennas can do.

How strong is the echo?

Sharpness is worthless if the echo is drowned in noise. A radar's whole performance collapses into one number: the signal-to-noise ratio (SNR). It gathers up every effect we have met — transmit power, antenna gain, wavelength, the target's reflectivity, the fourth-power range loss, the receiver's thermal noise — plus the processing gains radar earns back by being clever about its pulses. The result is a budget you can actually compute:

SNR = P_t G_t G_r λ² σ⁰ δ_az δ_rg · (τB) · N_p ⁄ [ (4π)³ R⁴ · kTB · L ]

Every symbol in that equation, in plain English

P_t

Transmit power — the watts the antenna radiates during a pulse.

G_t, G_r

Antenna gain, transmitting and receiving — how tightly the beam is focused. It is the same antenna doing both jobs, so its gain enters twice: the reason antenna area counts squared.

λ²

Wavelength squared — for a given gain, longer wavelengths mean a physically larger collecting aperture. Set by the band you chose in the last section.

σ⁰

The target's reflectivity per unit area, from the material explorer — a typical mixed ground pixel sits near −15 dB.

δ_az δ_rg

The pixel's ground area, azimuth times range. Multiplying it by σ⁰ turns "reflectivity per unit area" into the actual echoing patch — which is why finer pixels return less.

τB

The time-bandwidth product: pulse width times bandwidth, the gain earned by compressing the chirp of the previous section.

N_p

The number of pulses added coherently while the target stays in the beam — a processing gain we take on credit here and earn in Part 2.

(4π)³ R⁴

Geometric spreading, out and back — the fourth-power range loss from the start of this part.

kTB

Thermal noise power: Boltzmann's constant × system temperature × bandwidth. The receiver's noise figure inflates T — an imperfect amplifier hisses at the echo it is trying to hear.

L

Losses — cables, hardware imperfections, and the atmosphere (which is why the figure has a rain switch).

There is no need to work through it by hand — the interactive figure is this equation. It is the engine the rest of this essay runs on: later we will drop in the numbers for real satellites, and eventually for Starlink's own antennas, and simply read the SNR off this same panel. For now, explore. Sweep the bandwidth and watch SNR fall as detail improves — the two really do trade against each other. Drop from 500 km to 350 km and the two-way path loss eases, lifting the echo by several decibels. Switch bands, and turn on rain to see who suffers.

Engineers fold this whole budget into a single figure of merit: NESZ, the noise-equivalent sigma-zero — the σ⁰ of the faintest surface whose echo would exactly break even with the noise. Breaking even is a low bar. A 0 dB pixel is half noise — enough to detect that something is there, but speckle-dominated as imagery. Operational designs aim for noise floors around −20 dB, so that typical −15 dB ground sits well clear. Whenever a bar in this essay just barely crosses zero, remember that margin, not the crossing, is what makes a picture. Part 5 returns to this.

The equation has a few consequences worth pausing on. Doubling the bandwidth costs about 3 dB — and not for the obvious reason. The wider receiver does admit twice the noise, but pulse compression pays that back exactly; the real cost is that a finer range cell is a smaller patch of ground, and a smaller patch returns less energy — sharper pictures are dimmer pictures. Changing the pulse repetition frequency, meanwhile, doesn't move the SNR at all: at a fixed duty cycle, more pulses just means proportionally shorter ones, and the energy leaving the antenna each second is unchanged. And, hidden in the way the terms combine, a bigger antenna is enormously favorable — hold everything else fixed and the SNR climbs as the square of the antenna's area. Keep that square in mind: Starlink's largest antenna is very big indeed, and it will dominate everything in Part 5.

Is the model real?

A calculator is only worth trusting if it reproduces reality. Before we point it at Starlink, let's aim it at satellites whose numbers are public and see whether it agrees. But first, a word about frequency. Radar lives in named bands, and each is a different compromise between resolution, atmospheric penetration, and what the law allows. Bandwidth — the thing resolution actually needs — is not a property of the carrier, but wide slices are far easier to engineer, and to get allocated, high up: 600 MHz is a modest cut of X-band and more than half of everything L-band has. So the high bands are where sharp imagery lives, at the price of being chewed up by rain and water vapour; lower frequencies shrug off weather but need enormous antennas — and pay a different toll, in the ionosphere, which twists and delays long wavelengths enough that an L-band SAR like NISAR has to measure and subtract the effect to stay sharp.

There is a legal map under that chart, too. Transmitting radar from orbit is only lawful inside specific allocations — the Earth-exploration-satellite active bands, slices set aside by international agreement. Starlink's licences are nothing of the kind: they are fixed-satellite-service communications allocations, and blasting a radar waveform through them would be a licence violation that any regulator, or amateur with a dish, could detect. That will matter in Part 5, when we ask what a covert radar would actually take.

Now let's run the test. ICEYE flies the most numerous commercial radar satellites — small X-band craft with a modest antenna. NISAR is a NASA/ISRO giant with a 12-metre dish working at long L- and S-band wavelengths — and, flying since July 2025, its figures are no longer brochure promises but a year of on-orbit measurement. They sit at opposite corners of the design space, which makes them a stringent pair to reproduce. Feed each one's published parameters into our engine and compare what it predicts against what the operators report.

And they do match, closely. For ICEYE the engine predicts 3.0 m range resolution and a noise floor of −16.2 dB against a published 3 m and roughly −16.5 dB. For NISAR it predicts the same 3.0 m and a noise floor near −31 dB, comfortably inside the mission's ≤ −25 dB requirement. A disclosure: each satellite gets its own loss allowance (3 dB for ICEYE, 6 dB for NISAR's more complex chain), and ICEYE's 3.2 m × 0.4 m antenna is stood in for by an equivalent ~1.1 m square radiating 4 kW peak. This is calibration, not prediction — which is why two benchmarks from opposite corners of the design space beat one: the same few allowances have to satisfy both at once, and they do.

So we can trust the tool. But a link budget is a light meter, not a camera: it says how strong the echo is, not how millions of echoes become a picture. Before we can point anything at Starlink, we have to understand how a radar image gets built at all — a trick strange enough to deserve a part of its own.

Part 2Building a Picture

In Part 1 we learned how strong and how sharp a radar echo can be. But an echo is not a picture. A camera has millions of pixels; a radar antenna, pointed at a scene, sees just one. So how does a satellite build a crisp image from a single fat beam? That is the strange and beautiful trick of synthetic aperture radar (SAR), and it is what this part is about.

The one-pixel problem

The angular sharpness of any antenna is set by its size measured in wavelengths: roughly λ / D. A camera lens is tens of thousands of wavelengths of visible light across, so it resolves millions of angular pixels. A radar antenna is large by that measure — ICEYE's X-band array is about 3 m long, roughly a hundred wavelengths — and it is still nowhere near enough: a hundred wavelengths makes a beam about 0.6° wide, and from orbit a 0.6° beam paints a footprint kilometres across. Everything inside it echoes at once — a single blurry angular pixel. To make an image the old way, a radar would have to sweep that fat beam across the scene, one coarse pixel at a time.

Only one axis is blurry, though. Part 1's c / 2B slices the scene finely in range — hundreds of crisp distance bins inside that footprint, at any distance. What the radar cannot do is tell two echoes apart across the beam: one pixel across the beam, hundreds along the range. SAR's job is to invent sharpness in the one direction that lacks it.

How big an antenna would fix it? To resolve 1 m from 700 km away, the beam must be 1 ÷ 700,000 ≈ 1.4 millionths of a radian wide. At X-band's 3 cm wavelength, λ / D run in reverse demands D = 0.03 ÷ 0.0000014 — an antenna about 21 kilometres across. No one can fly a twenty-one-kilometre dish. Remarkably, though, there is a way to fake one.

Faking a kilometre-wide antenna

The insight that makes SAR possible is hiding in the satellite's own motion. As it flies, the same patch of ground stays in its beam for about a second — a kilometres-wide footprint crossed at 7.6 km/s — and at the kilohertz pulse rates radar uses, a second is thousands of pulses. If we record every echo along the way and combine them knowing exactly where the satellite was for each, the string of positions acts like one giant antenna — a synthetic aperture as long as the flight path over which the target stayed in view.

The engine that separates targets along the flight direction is the Doppler effect. A scatterer the radar is flying toward returns a slightly higher frequency; one it is flying away from, a slightly lower one. Since the shift is zero exactly broadside and grows as a target sits further forward or aft in the beam, Doppler is really a measurement of angle — and angle plus range pins a position on the ground. That spread of frequencies — the Doppler bandwidth — becomes the second axis of the image, called azimuth. Drag the satellite across the scene and watch the synthetic aperture build and the returns fan out in Doppler.

These are not two separate mechanisms. "The string of positions acts like one giant antenna" and "Doppler separates the targets" are the same fact told in two languages — one in geometry, the other in frequency. A big antenna resolves angle by comparing the phase of one wavefront across its width; the synthetic aperture compares the phase of successive echoes across the flight path. The phase differences are the same, and so are the angles they encode.

This leads to one of the most counterintuitive facts in radar: a smaller antenna gives better azimuth resolution. A small antenna has a wider beam, so a target stays in view longer, so the synthetic aperture is longer. In stripmap mode — the antenna staring fixed and sideways while the strip of ground scrolls past — the azimuth resolution works out to about half the physical antenna length — independent of range, just like the range resolution we met in Part 1. With one trick per axis, the one-pixel radar images like a camera.

That rule seems to contradict Part 1, which said a bigger antenna is enormously favorable. It is — for sensitivity, for swath, and (as Part 3 will show) for headroom against ambiguities; the D/2 rule says only that stripmap sharpness runs the other way. And big antennas have an escape hatch: spotlight mode, where the beam is steered to stare at one chosen spot as the satellite flies past, stretching the dwell and the synthetic aperture far beyond what a fixed beam allows — finer azimuth resolution, bought by giving up coverage of everything else. A designer picks which to spend. Starlink's antennas are metres across, not kilometres; whatever imaging they could do rests entirely on the synthetic aperture.

There is a catch: the radar must pulse fast enough to sample that Doppler bandwidth — the pulse repetition frequency must exceed it, or the image ghosts. That tension is the heart of Part 3.

Watching a picture come into focus

First, a word about what a radar sample actually is. Each echo sample is a complex number — picture a little arrow with a length and a direction. The length is the echo's strength; the angle is its phase, where in its cycle the wave was when it arrived. The radar's clock is steady enough that this angle survives the round trip to the ground and back, pulse after pulse — that is all the word coherent means: the radar remembers phase from one pulse to the next. Every trick from here on — focusing, interferometry, clutter cancellation — is arithmetic on those arrows.

The raw data a SAR collects looks like nothing at all — a smear of noise called the phase history, with fast time (range) across the rows and slow time (one row per pulse) down the columns. Turning that smear into an image is a sequence of signal- processing steps. The classic recipe is the range-Doppler algorithm, and the figure below walks through it stage by stage. Step through it and watch a blur resolve into a ship.

Two of these steps deserve a name. Pulse compression is a matched filter: it correlates the noisy return against the exact chirp that was transmitted, collapsing a long faint pulse into a sharp bright spike and lifting the signal out of the noise — the same trick that lets GPS work below the noise floor. Azimuth compression is a second matched filter, this time along the flight direction, that focuses each target's Doppler history to a point. In between, range-cell-migration correction straightens the gentle curve a target traces through the data as the slant range changes.

Why this recipe won is a matter of speed. Each matched filter is a correlation, and the fast Fourier transform turns correlation into simple multiplication — roughly N log N operations instead of comparing every pixel against every pulse. For a 16-million-pixel image built from a few thousand pulses, the brute-force count is around 10¹¹ operations; the FFT route is closer to 10⁹ — about a hundredfold cheaper. That hundredfold comes due two sections from now.

Go deeper: why SAR images look so strange

Because one axis of the image is range, not a viewing angle, SAR pictures carry artifacts a camera never shows: tall objects "lay over" toward the radar, radar shadows stretch away from it, and bright point scatterers ring with sidelobes. There is also the grain. Each resolution cell holds many small scatterers whose arrows add with random phases — sometimes reinforcing, sometimes cancelling — so even a perfectly uniform field comes out salt-and-pepper. This speckle is in every real SAR image; averaging several looks smooths it, at the price of resolution. The "30°–60°" grazing-angle rule has a reason at each end. Look straight down and every scatterer at a given range collapses into one bin, and the scene cannot be resolved at all; look too shallow and the opposite happens — shadows stretch until whole valleys vanish into black, and any slope tilted toward the radar compresses into a bright sliver.

When things move, they lie

The whole range-Doppler machine assumes the scene holds still for the seconds it takes to build the synthetic aperture. A moving target breaks that assumption, and it does so in a very specific, revealing way. Motion along the flight direction smears the target into a streak. Motion toward or away from the radar adds a false Doppler shift — and remember from the last section that Doppler is the azimuth axis, so the extra shift displaces the target along the flight direction, in azimuth. A moving car gets painted in the grass beside the road it is driving on.

The displacement is large. It works out to the target's radial speed times the range, divided by the satellite's speed: a car closing at 20 m/s, seen from 700 km by a satellite doing 7.6 km/s, lands 20 × 700,000 ÷ 7,600 ≈ 1,800 m out of place — painted nearly two kilometres from the road it is on.

This looks like a defect, but it is also an opportunity. The artifacts encode the target's velocity — if you can measure the azimuth displacement, you can recover how fast it was moving toward you. The obvious move — measure the offset and slide everything back — fails for a frustrating reason: in a single image, a displaced mover and a stationary object that really is sitting in the grass look identical. Breaking that tie takes a second measurement, and that is exactly why Part 3 will end up splitting the antenna into two. Turning this leakage into reliable detection and tracking of every moving vehicle is the entire subject of that part. First, though, one more way to form the image — the one that makes video possible.

Back-projection, and radar video

The range-Doppler algorithm is fast but rigid. A more brute-force method, time-domain back-projection, is slower but wonderfully flexible — and it is what makes moving-target tracking work. The idea is simple: lay down a grid of pixels on the ground, and for every pixel, for every pulse, look up how far the radar was from that pixel, grab the echo that arrived at exactly that delay, and add it in. A pixel with a real scatterer sees the same strong return align pulse after pulse and sums up bright; an empty pixel sees random phases that cancel.

The algorithm is famously compute-hungry, and now we can say how hungry: every pixel visits every pulse, so the 16-million-pixel image that cost the range-Doppler recipe about 10⁹ operations costs back-projection pixels-times-pulses — around 10¹¹, a hundred times more. It is why you rarely see live SAR back-projection. But a graphics processor (GPU) can run it for every pixel at once. The figure below is the real algorithm in a fragment shader: a hundred pulses, back-projected across the whole image in real time. Watch the aperture build and the point targets snap into focus; sharpen them with more bandwidth.

That flexibility does not repeal ambiguity, though: how often the radar pulses sets a hard sampling limit, and data undersampled in one domain is undersampled in every domain — no algorithm conjures the missing samples back (the next part opens with exactly that wall). What back-projection buys is freedom of geometry: it follows any trajectory, however curved, with no straightening approximations — and, best of all, you can point the same math at a moving grid to freeze a target that would otherwise smear. Split a long collection into overlapping intervals and you get a series of frames — video SAR. This is the engine behind custodial tracking: keeping a continuous radar eye on one jet, one ship, one missile. We will put it to work in the next part.

Nothing in back-projection cares whether the pulse was transmitted by the satellite that receives the echo — or by a different one entirely. That freedom will matter later, when transmitter and receiver end up on different spacecraft.

Part 3Tracking Motion

Set Starlink aside for a moment. To judge whether it could ever do this, we first need to know what "this" is — and nobody has wanted it longer than the US military. A still image misplaces everything that moves — which means a radar that can correctly place moving things can track them. Doing that over wide areas, for targets from walking soldiers to Mach-8 warheads, is the hardest problem in the field — a persistent, all-weather tracking capability the Pentagon has pursued for decades, and that defence reporting calls a "holy grail" of surveillance. The US is now building exactly this — a Space Force / NRO constellation of ground-moving-target radar satellites, with launches slated (as of 2024–25 reporting) from 2028. This part builds the tools, one tension at a time.

The Doppler dilemma

Everything hinges on one number: the pulse repetition frequency, or PRF — how often the radar shouts. It is squeezed from both sides. It must be high enough to sample the Doppler bandwidth, or moving and off-centre returns alias into ghosts. But it must be low enough that the echo from one pulse returns before the next goes out, or distant returns get eclipsed and ranges fold on top of each other. Better resolution widens the Doppler bandwidth and pushes the floor up; a wider swath pushes the ceiling down. Slide the PRF and watch the two walls close in.

Run the two numbers yourself to feel the squeeze. The Doppler floor: at 7.6 km/s, azimuth resolution δ demands a PRF of at least v/δ, so a 25 cm image needs 7,600 ÷ 0.25 ≈ 30,000 Hz. The range ceiling: the far edge of one pulse's swath must echo back before the near edge of the next, PRF ≤ c/(2·swath) — for even a modest 25 km swath, 300,000 ÷ 50 ≈ 6,000 Hz. A floor of thirty thousand does not fit under a ceiling of six thousand, and resolving that squeeze is what the rest of this part's machinery is for.

A wrinkle: real radars keep several pulses in flight at once — from 700 km the round trip takes nearly 5 ms, dozens of pulse intervals — so the ceiling is not literally one-pulse-at-a-time. It is about sorting echo windows: each pulse's returns must land in a window no other pulse's returns can reach. The squeeze is the same; the loopholes in it are where the next section lives.

Go deeper: one picture holds the whole squeeze

Range resolution, velocity resolution, and this PRF dilemma are not three facts — they are one, and there is a single picture that shows it: the ambiguity function, the radar's response spread over delay (which reads as range) and Doppler (which reads as velocity). Switch the waveform and watch it change shape.

The last panel is the dilemma in a single picture: the whole tension is the shape of one box that cannot be made large in both directions at once.

Slicing the sky

The first way out costs nothing but timing. A phased array can steer its receive beam far faster than echoes return, so it can scan a sharp pencil beam to catch only the instant of ground return — scan-on-receive. This masks returns from the wrong ranges (relaxing the range-ambiguity ceiling and raising the usable PRF) and adds receive gain for free. The idea is flying today: NISAR — the 12-metre giant that validated our engine in Part 1 — uses exactly this trick under the name SweepSAR, and it is how a single beam's worth of PRF budget stretches across a swath more than 240 km wide, several times what a fixed beam could support at full resolution. Push the same idea upward and you can time the scan to a chosen altitude, isolating aircraft at that height with no ground clutter at all.

The timing works out comfortably. At a 45° grazing angle, an aircraft at 10 km altitude sits about 7 km closer along the slant than the ground beneath it, so its echo arrives some 47 microseconds early — an eternity to a beam that can repoint in under a microsecond. Aim the receive scan at that early window and you image a shell of sky at 10 km: every echo in it flew, because nothing on the ground could have arrived that soon.

Slicing the sky handles aircraft. But most of what a military wants to track is on the ground, buried in a blaze of stationary clutter thousands of times brighter than the target. For that, one beam is not enough — you need several, and the difference between them.

Two eyes see motion

Split the antenna into two phase centres a short distance apart and you get two nearly-identical images. Now recall the arrows from Part 2: in each pixel, take one image's arrow and rotate it back by the other's angle — that is all "multiply by the complex conjugate" means — and the stationary world vanishes: whatever didn't move between the two looks cancels to zero phase, but a moving target is left standing at an angle, and that angle is a direct measure of its radial velocity. This is along-track interferometry.

How big is the angle? Between the two looks the mover closes some distance toward the radar, and the phase turns a full circle for every half-wavelength closed: Δφ = 4π·v_r·Δt/λ. Put in the essay's numbers — phase centres 4 m apart on a satellite at 7.6 km/s see the scene Δt = 4 ÷ (2 × 7,600) ≈ 260 microseconds apart — and at X-band's 3 cm wavelength a car closing at 20 m/s turns the phase by about 2.2 radians, near 130°: unmissable. Run it backwards and you get the minimum detectable velocity: if noise lets you trust the phase only to a few degrees, the slowest mover you can call real is somewhere under a metre per second. A person walking at 1.4 m/s clears that floor — but not by much, and a stroller or a drifting boat can sink below it.

A close cousin, displaced phase-center antenna (DPCA), spaces the two looks by exactly the distance the satellite travels in one pulse. Then the second antenna sees precisely the clutter the first saw a pulse earlier, and subtracting the two cancels the stationary world outright, leaving only movers. The catch is that across a wide swath the stationary ground is not one thing: points at different look angles close on the radar at different rates, so the geometry alone smears the clutter into a band of apparent velocity — and Earth's rotation shifts and broadens that band further. DPCA alone cannot cancel the whole smear edge to edge. (STAP, two sections on, fights a different beast — a sharp diagonal ridge in angle-Doppler space. What we have here is just a swath-wide smear.)

The same phase, read as a ruler

Along-track interferometry compared two looks taken a fraction of a second apart, and read motion out of the difference. Separate the two looks across the track instead — two antennas offset sideways, or the same satellite flying the same orbit twice — and the identical phase comparison measures something else entirely: not how the ground moved, but its shape.

Each point is a little closer to one antenna than the other, and that path difference wraps the phase into fringes. One fringe is a fixed step of elevation, so counting them from a valley to a summit rebuilds the terrain — a contour map drawn in radio waves. This is how one radar can measure the height of a whole planet: in 2000 the Shuttle flew two SAR antennas on a 60-metre mast and mapped nearly all of Earth's land in eleven days.

Now fly the same orbit twice, weeks apart, and subtract one pass from the other. Everything that held still cancels to nothing. Anything that crept toward or away from the satellite in between is left drawn in fringes — but now each fringe is not metres of height, it is half a wavelength of motion, a few centimetres. A city sinking as its aquifer is pumped dry, a volcano swelling before it erupts, the ground lurching in an earthquake: all of it comes back as bullseyes of colour, measured from three hundred kilometres up to better than a centimetre. This is the most widely used trick in all of civilian radar, and it runs on the exact phase Part 3 has been spending on movers.

There is a catch, and it is a second sensor in disguise. The comparison only means anything where the fine speckle pattern still matches between the two passes — where the looks stay coherent. Disturb the ground between passes — plough it, drive over it, dig it — and the speckle scrambles; coherence collapses. That collapse betrays a tyre track or a footpath that changed the surface without changing how bright it is: invisible to the amplitude image, glaring in the coherence one.

Topography, deformation, and the ghost of anything that moved the soil — all from comparing the phase of two looks. But for genuine movers, along-track interferometry left a problem open: a single baseline wraps, and a fast target reads as a slow one. Resolving that takes more than two antennas.

Resolving the velocity ambiguity

Interferometry measures velocity as a phase — and phase wraps. Beyond a maximum radial velocity, a fast target's phase rolls past a full turn and reads as something slower; at just the wrong speed it wraps all the way to zero and vanishes entirely (a blind speed). A single baseline forces a cruel trade: make it long to detect slow movers, and fast movers alias; make it short for fast movers, and you lose the slow ones.

The escape is centuries old: use several baselines and the Chinese Remainder Theorem. Each baseline wraps at a different velocity, so the true velocity is the one value consistent with all of their phase readings at once. The theorem itself fits in three lines: a whole number that leaves remainder 2 when divided by 3, and remainder 3 when divided by 5, can only be 8 — the next candidate sits a full 3 × 5 = 15 away. Now read "remainder" as a wrapped phase reading and 3 and 5 as the speeds at which two baselines wrap: the true velocity is the one value both readings agree on, and the ambiguity is pushed out to the product of the wrap speeds, not their sum. Set a target speed and the baselines below, and watch the ambiguous candidates collapse to a single answer.

This is not exotic: ordinary weather and air-traffic radars have played the same trick for decades, alternating between a handful of pulse rates and keeping the velocity consistent with all of them.

With a well-chosen set of baselines you can, in principle, read velocities unambiguously from a walking pace to well past the speed of sound — though that sentence hides two IOUs: the baselines must all fit on one spacecraft, and phase noise can shove a reading across a wrap boundary and snap the answer to a wildly wrong value, which is why "robust" Chinese-remainder unwrapping is a research field of its own. And interferometry still struggles when the signal is weak and the clutter broad. For the hardest cases — very slow targets over very wide scenes — you need the heavy machinery.

Space-time adaptive processing

The most powerful clutter-suppression technique treats space and time together. With three or more phase centres, each pulse-and-antenna pair becomes a dimension in a data cube. Stationary clutter is highly correlated across adjacent antennas and adjacent pulses, so it collects along a tight diagonal clutter ridge. Space-time adaptive processing (STAP) learns that ridge from the surrounding data, then builds a filter that nulls exactly those correlations while passing everything else — digging out targets that were buried far inside the clutter.

STAP is compute-hungry — it inverts a covariance matrix per range cell — but for the hardest case, something person-slow somewhere in a wide, Earth-rotation-smeared scene, it is the closest thing the field has to a reliable answer. Its real-world catch is that it must learn the ridge before it can null it, and it learns from training data: the range cells surrounding the one under test. The rule of thumb is brutal — several times more clean, statistically-alike training cells than adaptive degrees of freedom — and real terrain is not statistically alike: fields give way to towns, roads, shorelines, and the estimate sours. That is the problem "knowledge-aided" variants exist to solve: seed the filter with terrain maps and known road networks, so it starts near the answer instead of learning the world from scratch — priors a giant always-on constellation could supply for free. It behaves less like a camera and more like a wide-area early-warning radar, running alongside the imaging modes.

Go deeper: the covariance null

For a cell under test, the surrounding range cells are taken as clutter-plus-noise training data. Their covariance matrix captures how clutter correlates across every pulse-antenna pair; its inverse, applied to a bank of velocity-and-angle steering vectors, yields weights that maximise the signal-to-interference-plus-noise ratio for each candidate target while suppressing the ridge. A simple threshold on the filtered output then declares detections.

There is a pattern in this part. Every step wanted more phase centres: two for interferometry, several baselines for the velocity unwrap, more still for STAP's degrees of freedom. A single spacecraft runs out of room for them fast.

The holy grail: a distributed constellation

Every technique so far lives on one satellite. The truly radical idea separates the roles: one transmitter illuminates a scene, and a swarm of cheap receiver satellites — no high-power transmitter, just a big antenna and a laser link — all listen at once. On the order of a few hundred transmitters is enough to keep the whole Earth illuminated; after that, receivers are almost free, and each one adds a new baseline and a new viewing angle.

The payoffs accumulate quickly: a dozen or two satellites can watch one target at once, baselines in two directions enable true 3-D imaging (circular and tomographic SAR), and stealth aircraft, shaped to bounce energy away from a monostatic radar, light up for a receiver sitting off to the side.

None of this is free, and the price is measured in centimetres and nanoseconds. Coherent bistatic imaging demands that transmitter and receiver agree on where they are and what time it is to a small fraction of a wavelength — centimetre-level ephemerides, clocks aligned to a fraction of a nanosecond, held for the whole aperture. Starlink already flies most of that kit: GPS receivers on every spacecraft, precise orbit determination, and optical inter-satellite links that could, in principle, tie clocks across the fleet. What it lacks is radar-grade oscillator discipline and the cross-calibration to prove it — no new physics required, but hard engineering.

And the mode that ties the whole part together is video. Recall spotlight from Part 2 — steer the beam to stare at one spot and the dwell, and the resolution, stretches as long as you like. Point a swarm's worth of spotlight stares at one target, form the frames with back-projection — which never cared whose pulse it was — and you get custodial tracking: a continuous eye on any single object.

Listening, and the trouble with jamming

A radar this sensitive is also a superb listener. Turn off the transmitter and the same antennas become a signals-intelligence sensor, picking up any emitter in their band. Form thousands of receive beams at once, sweep them across a scene, and the pattern of received power pins down where each emitter sits — multi-beam source localization. From orbit, this could map every GPS jammer around an airport.

To see why the listening matters, consider jamming, which is normally a good bet against a radar. Part 1 taught the asymmetry: the radar's echo pays the fourth-power round trip, but a jammer's noise travels one way and pays only R² — watt for watt, the jammer arrives with an enormous power head start, which is why jamming has bedevilled radar since its invention. Against a constellation, though, the bet turns. To jam you must broadcast, and broadcasting confesses your position to a sensor that can switch to listening and geolocate you in moments. The array can then steer a null onto you — at a real cost: every null spends receive degrees of freedom, and nulling fails when the jammer sits close in angle to the thing being imaged. Against one satellite that trade can favour the jammer; against dozens of receivers at dozens of angles it falls apart, because no jammer can sit close to all of them and there are too many to blind at once. The contest is not hopeless for the jammer, but it is one the constellation is built to win.

Scan-on-receive, interferometry, multi-baseline unwrapping, STAP, distributed apertures, passive listening: every one was conceived for purpose-built military satellites that cost billions and fly in single digits. And the ingredient list they share — phased arrays, precise clocks and orbits, inter-satellite links, and above all numbers, many spacecraft sharing the sky — contains nothing Starlink doesn't already fly. Whether what it flies is good enough is a question about radios. Part 4 takes Starlink's radios apart, and Part 5 puts the question to them directly.

Part 4How Starlink Talks

Starlink was built to move internet, not to make radar images. But radar and communications are the same physics wearing different clothes, and the radios Starlink flies are extraordinary. This part is a tour of that hardware — modulation, antennas, amplifiers — so that in Part 5 we can ask what it could do if pointed at the ground as a radar.

Writing bits onto a wave

To send ones and zeros through space, you stamp them onto an oscillating carrier by nudging its amplitude and phase. Plot each symbol as a point whose distance from the origin is amplitude and whose angle is phase, and you get a constellation diagram. That angle is the same phase that focuses SAR images, cancels clutter, and steers beams — the same piece of physics doing a fourth job. Two points (BPSK) carry one bit each; four (QPSK) carry two; Starlink's 256-QAM packs eight bits into 256 points — known not because SpaceX said so, but because Humphreys and colleagues reverse-engineered the Ku-band downlink off the air: OFDM channels about 240 MHz wide, each carved into 1,024 subcarriers. More points mean more data — but they crowd closer together, so noise starts flipping bits. Add noise and watch the symbols smear.

So there is an optimal modulation for every signal-to-noise ratio: use the densest constellation whose clouds still stay apart. A modern system re-chooses it, per user, many times a second — and layers forward error correction on top, spending a fraction of the bits on redundancy so the link can run just below the raw bit-error cliff without falling off it.

Behind that adaptation is a hard law — the communications twin of Part 1's radar link budget. Shannon's channel capacity,

C = B log₂(1 + SNR)

caps the bits per second any link can carry at bandwidth B and a given signal-to-noise ratio — the same two currencies the radar equation trades in. Adaptive modulation is simply a system climbing as close to that ceiling as its constellation clouds allow. No cleverness gets above it; the only ways up are more bandwidth or more SNR, which is exactly the bargain radar strikes too.

Symbols, bandwidth, and the matched filter

How much spectrum does a signal need? Any waveform can be built from sine waves, and the Fourier transform is the prism that splits it back into them. A single digital symbol is a little rectangular pulse, and the Fourier transform of a rectangle is a sinc — so the faster you send symbols, the wider the band they occupy. In the ideal case the symbol rate simply equals the bandwidth. Drag the symbol rate and watch the spectrum breathe.

This same Fourier thinking powers the radar trick from Part 2. A chirp spreads its energy across a wide band and a long time; a matched filter correlates the echo against that exact chirp and collapses it into a sharp spike, lifting a signal from beneath the noise. The (τB) term parked in Part 1's SNR equation is exactly this processing gain — the time-bandwidth product. A 20 µs chirp swept across 300 MHz compresses 6,000-fold, and 6,000× is about 38 dB of signal-to-noise earned by waveform design rather than watts.

Communications has the same trick under different names, and they are worth keeping straight: spread spectrum is the transmitter's half — deliberately smearing a signal across far more bandwidth than its data needs by stamping it with a fast code — while matched filtering is the receiver's half, correlating against the known pattern to snap the smeared energy back into a spike; radar's chirp compression is one instance of that receiver-side trick. Together they are how a 25-watt GPS satellite reaches a phone from twenty thousand kilometres away — the same below-the-noise-floor trick we met in Part 2's processing chain.

Sharing the sky

One satellite serves thousands of users, so it must divide its spectrum among them. It can split time (take turns), split frequency (give each a channel), or — the trick that makes Starlink scale — split space. With space-division multiple access, tightly focused beams point at different patches of ground, so the very same frequencies can be reused over and over as long as the beams do not overlap. Drag the users and watch beams reuse the band.

The arithmetic is what makes space division impressive. Give a satellite N spot beams whose footprints do not overlap and it can pour the same spectrum into every one of them — the same hertz sold N times over. A Starlink satellite runs thousands of beams, so a band that would serve one cell serves thousands at once. Part 5 will come back for those beams.

Starlink also uses OFDM, slicing each channel into about a thousand narrow sub-carriers, which resists interference and — as we will see — dodges a problem that would otherwise cripple a wide-band phased array. That problem is the subject of the next section, and it is the single biggest obstacle to reusing these antennas as radar.

Steering a beam with phase

Every Starlink satellite steers its beams electronically with a phased array, and the mechanism fits in a sentence: every element is a tiny transmitter radiating the same wave, and the array shines in whichever direction the crests happen to line up. Fire every element together and they line up straight ahead. Phase-advance each element a little more than its neighbour and the line of reinforcement tilts — the beam steers, with nothing moving at all. Watch it happen: slide from one element to many and a shapeless ripple organises itself into a beam, then steer it.

Averaged over time, that interference pattern becomes the array's beam pattern — the polar plot below. One spacing rule governs the layout: keep the elements within half a wavelength of each other. Space them wider and the maths admits a second angle where all the waves also line up — a grating lobe, a false beam pointing somewhere you never asked. Drag the steering and watch the pattern respond — and note how energy leaks into grating lobes when the elements are spaced wrong or the array is driven off its design frequency.

That flat polar plot is really a slice through a three-dimensional shape. The beam is a lobe — a narrow finger of energy pointing where you steer it, wrapped in smaller side lobes. Drag it around.

Beam squint: the wideband curse

There is a catch, though, and it shapes everything that follows. A phase shift is not a true time delay — it steers one frequency correctly and every other frequency to a slightly different angle. A radar-grade wideband signal contains a whole span of frequencies, so a phase-steered array smears them across a range of directions — beam squint. The bigger and higher-gain the array, the narrower its beam and the more damage squint does. Run the numbers for Starlink: a Ku-band comms channel is about 240 MHz on a ~12 GHz carrier — two percent fractional bandwidth — and the squint is manageable. Radar sharp enough for 25 cm detail wants about 600 MHz, five percent, where a purely phase-steered array smears badly anywhere off boresight. The gap between a comms array and a radar array is that factor of two and a half. Widen the bandwidth and steer off-boresight to watch the beam fall apart by colour.

There are two ways past the wall, and Starlink's own signal already carries one of them — the dodge the last section promised. Because OFDM splits the signal into about a thousand narrow subcarriers, each subcarrier is effectively narrowband on its own; apply a slightly different digital phase to each one and the squint largely corrects itself, subcarrier by subcarrier. The brute-force alternative is a hybrid array: group elements into tiles, give each tile a true time delay (which is frequency-flat), and let analog phase shifters do the fine steering within a tile. Fewer elements per tile means less squint — at the cost of gain and complexity. Slide the tile size to trade one against the other.

A quarter-turn of polarization

One more knob per element: the direction the electric field oscillates, its polarization. Feed an element from two perpendicular directions with a phase shift between them and the field rotates — circular polarization — which talks cleanly to tumbling phones and, as Part 5 will use, offers a radar real isolation: transmit one handedness and receive the other, and you keep the single-bounce echo (which flips on reflection) while rejecting your own leakage (which doesn't). The price is that double-bounce and depolarized returns are rejected too, and the ~30 dB figure is the optimistic end for a large steered array.

Polarization is also a second channel of information, not just a duplexing trick. Transmit and receive on both horizontal and vertical and you measure not merely how bright a surface is, but how it bounced. One clean bounce off open ground keeps the polarization it arrived with — surface scattering. The right-angle of a wall meeting the ground bounces twice and flips it — double-bounce, the signature of buildings and ships. The tangle of a forest canopy scrambles it completely — volume scattering. Colour a scene by which mechanism dominates and a uniform grey image sorts itself by material: soil, city, forest, water.

This is why the fully polarimetric radar an operator would want can tell a tank from a truck from a hedgerow — and why imaging several bands and polarizations at once, as Part 5 will note, turns a grey picture into something closer to a material map.

From bits to radiated power: the RF chain

Behind every antenna element is a radio-frequency chain. Transmitting, it turns digital samples into an analog wave, up-converts to the carrier frequency, and amplifies; receiving, it runs the same steps in reverse, with a low-noise amplifier coddling an impossibly faint signal before an analog-to-digital converter hands it to the processor. Modern arrays push more of this into software-defined radio on a chip — flexible, but power-hungry.

The receive side sets limits of its own. The low-noise amplifier's noise figure feeds straight into the kTB noise floor of Part 1's SNR equation — a decibel lost there costs a decibel of transmit power everywhere else. And the analog-to-digital converter has finite dynamic range: it must digitise a faint echo without being blinded by whatever loud signal arrives at the same instant. That is the real "hear a whisper next to a scream" constraint, and the one Part 5's isolation ladder exists to solve.

The amplifier's compromise

The power amplifier forces the compromise that decides what Starlink could do as a radar. A transistor is a valve, faithful only in its middle range: drive it too hard and the peaks clip; let it idle in the linear zone and it wastes power as heat.

How you bias that valve sets the amplifier's class — the trade of linearity for efficiency.

What decides how hard you must back off is the peak-to-average power ratio. A clean radar chirp barely varies — its peak-to-average ratio is essentially 0 dB — while OFDM's violent peaks run 8 to 12 dB above the average, forcing the amplifiers to idle far below their limits. Hand those same amplifiers a chirp and they can run roughly ten times harder without clipping. That is the smaller of two levers. The bigger one is beam count: a comms array normally splits its total radio-frequency capacity across thousands of simultaneous beams, each serving a user at low power. Concentrate the entire array into one beam for a brief burst and the per-beam power scales by the beam count — the two levers together are worth orders of magnitude, for seconds at a time. The limit is not the transistors: bursts draw down batteries and dump waste heat faster than the radiators can shed it, so the duty cycle — how often the radar can afford to speak up — is capped by the power system, not by the physics. Part 5 will lean on this fact.

That is the machinery: waveforms matched to signal-to-noise, spectrum shared by beams, arrays steered with phase, amplifiers biased against their own physics — every piece designed to sell internet, and every piece something a radar engineer would recognise on sight. One section remains in Part 4: the actual antennas flying today. Then Part 5 asks whether they can be turned around.

The antennas Starlink actually flies

With the machinery understood, we can meet the actual hardware. A V2 Mini carries four downlink antennas; a V3 carries five. Ku-band is the workhorse user downlink, the beam that reaches the pizza-box terminals. Ka-band and E-band are gateway links — fat pipes to ground stations, not to customers — with E-band, at 75 GHz, carrying the heaviest backhaul. V3 adds a V-band antenna that V2 never carried: new spectrum, not a grown version of anything. And then there is direct-to-cell, near 2 GHz, whose job — reaching an ordinary phone with no dish at all — forces it to be huge.

Antenna	Frequency	Role	Aperture	Gain	TX power (operational)
V2 Mini — 500 km
E-band	75 GHz	gateway feeder	~0.5 m	52 dB	200 W
Ka-band	30 GHz	gateway link	~0.7 m	47 dB	260 W
Ku-band	11.7 GHz	user downlink	~1.3 m	44 dB	27 W
Direct-to-cell	2 GHz	ordinary phones	~3.8 m	38 dB	175 W
V3 — 350 km
E-band	75 GHz	gateway feeder	~1.4 m	61 dB	98 W
V-band	40 GHz	gateway / capacity (new on V3)	~1.3 m	55 dB	66 W
Ka-band	30 GHz	gateway link	~1.0 m	50 dB	120 W
Ku-band	11.7 GHz	user downlink	~3.2 m	52 dB	165 W
Direct-to-cell	2 GHz	ordinary phones	~15 m	50 dB	15 W

Gains and powers are the model's working figures, assembled from filings and public reporting; the apertures are the sizes those gains imply at each frequency.

Read down the table and two things stand out. First, every one of these — the millimetre-wave gateway links included — is a phased array pointed at Earth: no dishes, no gimbals, just flat panels steering beams with phase, exactly the machinery of the last three sections. Second, one row does not fit the pattern. The gateway and user antennas are all metre-class or smaller; the direct-to-cell array is a few metres across on V2 Mini and roughly fifteen on V3, because physics offers no other way to close a link with an unmodified phone. Its gain says 50 dB; its area says something the gain column understates, since Part 1 showed SNR climbing as the square of antenna area.

That is the full inventory: nine antennas across two generations, each with a band, an aperture, a gain, and a power budget. Part 5 will feed exactly these numbers into the link budget from Part 1 and ask each antenna, in turn, whether it could image the ground.

Part 5Is Starlink a Secret Radar?

We now have every piece. Part 1 gave us the radar link budget; Part 2, how an image is built; Part 3, how motion is tracked; Part 4, the specific radios Starlink flies. This final part puts them together and asks the title question: could Starlink's existing antennas — designed to sell internet — work as a radar and signals-intelligence constellation? Two sections from now, we will run the numbers, antenna by antenna.

The idea of one aperture that both talks and senses is not a fringe notion. Engineers call it integrated sensing and communication, and it is one of the pillars people are designing 6G around. Starlink would be the largest instance ever flown — whether or not anyone set out to build one.

Three problems stand between a communications satellite and a radar. It must hear a whisper while shouting. Its comms hardware must be coaxed into radar-like behaviour. And even if the physics closes, the torrent of data a radar produces — and the compute to digest it — has to go somewhere. We take the first problem first.

Hearing a whisper next to a scream

Radar has a cruel problem built into it: the pulse you transmit is overwhelmingly louder than the echo you hope to hear. Part 1's own budget puts numbers on it. A radar-grade burst from orbit is tens of kilowatts, about +43 dBW; the raw echo returning from a single −15 dB patch of ground, before any processing gain, arrives near 10⁻¹⁵ watts, about −150 dBW. The gap between those two numbers is nearly two hundred decibels — a factor of ten billion billion. Detecting the echo while shouting is like hearing a pin drop during a rocket launch. The art of managing this is duplexing, and it comes down to stacking up isolation from every trick available.

One row on that ladder deserves special attention: spatial separation. Put the transmitter and receiver on different satellites and distance does the heavy lifting: at a hundred kilometres of separation, the direct path near 2 GHz already loses about 138 dB in free space — the single biggest bite anyone takes out of the ~200 dB requirement, leaving some 50 to 60 dB for polarization, beam nulling, and active cancellation to mop up. This "bistatic" arrangement — one Starlink illuminating, another listening — is the key that unlocks the most powerful modes, and its price is the one Part 3 already put on the table: the two spacecraft must share a clock and know their baseline to a fraction of a wavelength.

None of this is hypothetical, either. Researchers have already used Starlink's own downlink as the transmitter of a bistatic radar: the first published signal measurements of Starlink as a radar illuminator appeared in 2022, and experimental images of the Earth's surface formed from Starlink downlink echoes have followed. The satellites were never asked — all the listening happened on the ground. The premise of this part has already been demonstrated in practice.

Why does this matter so much? Because a radar that can listen and transmit at the same time — full duplex — is freed from pulsing altogether. A normal radar must go quiet to listen, which caps how long and how often it can transmit. Lift that limit and it can run a continuous wave, pour in far more energy, sample motion far faster, and see a much wider swath at once. The figure below shows what the constraint costs, and what removing it buys.

Can each antenna form an image?

We now have everything we need — a validated calculator, and an understanding of resolution, power, beamforming, and duplexing — so we can finally take every antenna Starlink flies and ask, one by one, whether it could form a radar image. The model is exactly what you would build by hand from the table that closed Part 4: each antenna enters with its aperture, its gain, its bandwidth, and two power levels — everyday operation and a radar-optimised burst — with the V2 Mini fleet at 500 km and V3 at 350 km. Every bar is an antenna you have already met. The verdict is a single number — is the SNR above zero for a realistic ground target? Green means yes.

The chart flies V2 Mini at 500 km and V3 at 350 km. Hundreds of V2 Mini direct-to-cell satellites already orbit in ~350 km shells, so part of the V2-to-V3 improvement you see is an altitude assumption — several decibels of fourth-power range loss — and not hardware at all.

The current V2 Mini fleet is nearly all red. Its high-frequency phased arrays have too little power for their bandwidth; even the Ku-band antenna, driven in a burst, stays a few decibels underwater. (One bar is already green — we will come back to it.) And whatever the bars say, doing radar covertly with this fleet is impossible: a radar burst in the Ku or Ka bands would be a loud, out-of-allocation transmission in spectrum every regulator, defence monitor, and radio amateur can watch, and flagging exactly that kind of anomaly is what they do. If the current fleet were imaging the ground, someone would have heard it by now.

The picture changes with Version 3. Its Ku, Ka, and E-band antennas all grow in size and gain, it adds a V-band antenna V2 never carried, and the whole spacecraft drops to a lower orbit. Not all of these face customers — E-band is a gateway feeder and Ka mostly serves gateways too — though which customer a beam was sold to makes no difference to the physics. Now the Ku-band antenna is a capable imager, the higher bands hover at the edge of usefulness in clear skies, and — because each band reflects differently off each material — imaging several at once would yield false-colour radar that could tell aluminium from steel from flesh.

Part 4 left an objection standing: beam squint. Radar-grade bandwidth is a few percent of the carrier for most of these antennas — 500 MHz on the Ku carrier is about four percent — and a purely phase-steered array smears exactly such signals when steered far off boresight. The outs are the two from Part 4: OFDM-style processing that applies its own digital phase per narrow subcarrier, and true-time-delay tiling in the hardware. Starlink's arrays already run versions of both for communications; radar would push them harder — hardest of all on the direct-to-cell antenna, where 500 MHz is a full quarter of the 2 GHz carrier. The bars assume the steering problem is solved — and it is solvable, with machinery Starlink already flies, though not for free.

The bars also judge success as "SNR above zero", the permissive floor Part 1 flagged; real designs want 10–20 dB of margin. Hold the chart to +10 dB and only V3's Ku-band antenna and the direct-to-cell antennas stay green; demand 20 dB and the direct-to-cell antennas stand alone.

Notice that every claim in the covertness paragraph rested on the same premise: a hidden radar fails because it must transmit loudly, and loud transmissions are visible. The V3 chart holds one exception to that premise — an antenna that reaches a usable image while radiating little more than a phone call. The next section looks at it closely.

The quiet one: Direct-to-Cell

The V3 direct-to-cell antenna exists to connect ordinary phones to satellites, so it is enormous — about fifteen metres across — and it works around 2 GHz, down in the S-band of Part 1's chart, at wavelengths that sail through weather and even through the plastic shells of drones. Remember that a big antenna helps as the square of its area — and by the model's gain figures this one has sixteen times the area of its predecessor. The consequence takes a moment to sink in: this antenna reaches a usable image while radiating just a handful of watts.

That resolution number needs care, though: the readout is range resolution, the across-the-swath axis that bandwidth buys. The other axis obeys Part 2's rule — stripmap azimuth resolution is about half the antenna's length, and half of fifteen metres makes the pixel eight metres long. That is the price of the giant antenna: its sensitivity is bought by coarsening azimuth, so the cheap pixel is half a metre by eight metres — sharp across the swath and blunt along it. Part 2 also described the remedy: spotlight mode. Stare at one scene, synthesise a longer aperture, and the azimuth sharpens toward the range figure — paid for, as always, in coverage.

Driven at full power, the numbers grow stranger still. The antenna's thousands of beams cannot all transmit at once — simultaneous beams divide the burst power, and per-beam SNR falls with the count, a trade the figure's model does not include. But bursts in sequence do the same work on a schedule: hop the beam from scene to scene and this one antenna could sweep vast areas at ten-metre-class resolution, its revisit rate capped by tasking, heat, and data rather than physics. More practically, the spotlight video-SAR tracking of Part 3 lets it lock onto a single moving target and film it: a fighter jet, a cruise missile, or a small drone whose plastic shell is transparent at these wavelengths but whose motors and battery are not. Even stealth is only a partial answer here. Shaping defeats a single viewpoint, as Part 3 explained — but a low-flying aircraft cannot stop itself from blocking the radar's illumination of the ground behind it, and that moving hole in the image is a track. The bistatic version of that idea is not hypothetical: a 2024 field trial used a Starlink satellite's own signal to catch a stealth-proxy drone by the disturbance it cast. Every space-time-adaptive and multi-beam trick we built earlier applies — this one antenna could run them all.

This is one antenna, on a satellite designed to sell phone service. As of August 2026, hundreds of V2 Mini direct-to-cell satellites are already on orbit with metre-class arrays, and the fifteen-metre class has begun to fly: the first twenty V3 satellites deployed from Starship's thirteenth flight in July 2026, with the constellation building through 2027. A single antenna, then, can see. Whether a fleet of them amounts to surveillance depends on how much of the Earth they cover, and how often — which is where we turn next.

How much, and how often

One antenna can form an image. But a single photograph of a single strip, taken once, is not surveillance. Two numbers decide whether a fleet watches the world: how wide a strip each pass paints — the swath — and how long a given place waits between looks — the revisit.

A satellite in low orbit never stays over one place. At 550 km it laps the Earth in about ninety-five minutes at 7.6 km/s, and while it loops the planet turns beneath it, so each orbit's ground track falls some 2,600 km west of the one before. A single satellite with a few-hundred- kilometre swath therefore paints thin diagonal ribbons with wide gaps between them. A particular town can wait many hours — sometimes a day — for a pass that comes close enough to image it.

This is why the capability is a constellation problem, not an antenna problem. Add satellites in orbital planes fanned around the Earth and their ribbons interleave; the gaps close. In the figure, one satellite leaves the marker waiting the better part of a day. Raise the count and watch that gap fall — a few satellites bring it to hours, a few dozen to minutes, and Starlink flies more than eleven thousand.

As always, nothing comes free: a wider swath covers faster but, from Part 2, blunts resolution or demands more power, and revisit is bought satellite by satellite. Still, persistent, near-real-time coverage of the entire planet is exactly what a swarm of steerable, wide-swath antennas delivers — which is the plain-language definition of the surveillance holy grail Part 3 opened with. The earlier parts settled the physics of the single look; the scale, as the first figure of this essay showed, is already overhead.

Why space datacenters — and what it all means

There is a bottleneck hiding behind every capability in this essay: data. Run the chain once: 500 MHz of bandwidth means five hundred million complex samples every second; at a few bytes per sample that is roughly two gigabytes per second of raw phase history; so a single wide swath, recorded for a few tens of seconds, is a hundred gigabytes. Forming an image of the whole Earth would take tens to hundreds of exaflops — compute comparable to the largest datacenters on Earth, applied continuously — and far more data than even laser links can carry to the ground. The only way to use this firehose is to process it where it is collected: in orbit.

Which suggests one reading of those announced orbital datacenters — an inference, not a fact. Serving chatbots from orbit will not out-compete ground-based GPU farms. But as the brain that turns a planet-scale radar and signals-intelligence firehose into finished intelligence, they make sense — and the customers for that capability have budgets measured in the hundreds of billions.

Final words

We started with a simple echo and built, step by step, to an uncomfortable conclusion. Nothing in this essay requires exotic new physics; it is the same radar equation we met at the beginning, applied to hardware that is already flying or announced. Starlink version 3 has the latent potential to be the most capable imaging and signals-intelligence constellation ever built.

"Latent" deserves a scorecard, because what separates potential from capability is a short list. Wideband steering — true time delay, or digital phase per OFDM subcarrier — is hard, and it is on the public phased-array roadmap. Inter-satellite phase synchronisation at radar grade is hard; the laser links already flying are most of the answer. On-orbit processing at the needed scale has been announced outright. And spectrum authority — a licence to emit radar waveforms, or the willingness to do it anyway — is not an engineering item at all; it is the tell to watch for. What makes the list unsettling is not its length, but how much of it already appears on the public roadmap.

Little of it is hypothetical any more, either. SpaceX already builds reconnaissance satellites for the National Reconnaissance Office on Starlink-derived buses, which reporting says carry synthetic-aperture radar among their sensors. The Space Force has a funded ground-moving-target program — a billion dollars in the 2026 budget, first launch slated for 2028. Independent labs have formed images of the ground from Starlink's own downlink. A field trial used a Starlink satellite to light up a stealth-proxy drone. And in early 2026 SpaceX filed to put datacenters in orbit. The pieces this essay assembled out of physics are, one by one, being built. In ninety years radar went from a secret ring of coastal towers to tens of thousands of satellites overhead; the last few steps — decoding Starlink's downlink, imaging the ground with it — have already been taken.

One question remains: what would give it away? Radar bursts in comms bands would be out-of-allocation emissions that any spectrum monitor — state or amateur — could flag; radio amateurs already fingerprint the signals of Starshield, SpaceX's classified constellation for the National Reconnaissance Office. A chirp looks nothing like OFDM to anyone who records one. Burst duty cycles leave thermal and power signatures. Licence filings are public. A planet that knows what to listen for is hard to image in secret.

A capability like this in the hands of any single actor — private or government — raises hard questions about surveillance, warfare, and the concentration of power, and about the legal loopholes that let commercial imagery be sold back to states that could not collect it directly. The physics behind it, though, is knowable, and an informed public is in a far better position to decide what should and shouldn't be built. Everything in this essay is now something you can check for yourself: scroll back up, move the sliders, and run the numbers.

Sources & further reading

This essay is an educational retelling of the video "Is Starlink A Secret Radar Constellation?" by Noise In Space, whose open-source simulations the physics here is faithfully ported from. Full credit for the original analysis belongs to the creator; the references below are where to go deeper.

• "Is Starlink A Secret Radar Constellation?" — Noise In Space: the original video this piece retells.
• NIS-Starlink-Radar-Video: the creator's open-source radar simulations (MIT License), from which every equation in this essay is ported.
• The War Zone — tracking ground and air targets from space: the defence reporting behind the "holy grail" surveillance claim in Part 3.
• Humphreys et al., "Signal Structure of the Starlink Ku-Band Downlink": the reverse-engineered downlink — OFDM, ~240 MHz channels, 1,024 subcarriers — behind Part 4's numbers.
• Starlink as a passive-radar illuminator (IEEE, 2022): the preliminary study and first signal measurements using Starlink downlinks for bistatic radar.
• First imaging results with Starlink illumination (IEEE): experimental Earth-surface images formed from Starlink downlink echoes — the demonstration cited in Part 5.
• Passive radar on LEO broadband satellites (IET, 2024): the survey of using Starlink and OneWeb as radar illuminators of opportunity — achievable performance, measured signal properties, and the first receiver prototypes.
• NPR — Starshield's signals: reporting on SpaceX's classified NRO constellation and the radio amateurs fingerprinting it.
• The War Zone — SpaceX's secret constellation: analysis of what Starshield may be and why it would matter.
• NRO Proliferated Architecture (Wikipedia): the public record of the classified constellation's launches.
• Tomographic bistatic SAR with the Starlink waveform: an end-to-end demonstration of passive 3-D imaging from Starlink signals, validated on real captures.
• Integrated sensing and communication over LEO constellations: the research field — one aperture that both talks and senses — that this essay is, in effect, an instance of.
• Interferometric SAR (Wikipedia): how comparing the phase of two looks measures terrain height and centimetre-scale ground deformation, the subject of Part 3's interferometry section.
• Synthetic-aperture radar (Wikipedia): how a moving antenna synthesises a large aperture for fine azimuth resolution.
• Radar (Wikipedia): fundamentals and the radar range equation that the SNR budget is built on.
• ITU-R Recommendation P.676: the standard model for atmospheric gas attenuation used in the band chart.
• ESA Sentinel-1 mission (SentiWiki): a canonical, well-documented operational C-band SAR constellation.
• NASA–ISRO NISAR mission: the L/S-band SAR giant, operational since 2025, used as a validation benchmark in Part 1.
• ICEYE: the commercial X-band SAR constellation used as the other validation benchmark.

The ported physics and simulations are © 2026 noiseinspacechannel, released under the MIT License (see the project's notices for full attribution). This retelling is an independent educational work; all credit for the original analysis belongs to Noise In Space.