Fundamentals

Why GPS does not work underground

What actually happens to the satellite signal in a metro station, why your phone keeps showing a false position, and what replaces it down there.

Published on 16 June 2026 Metrosafe team

Anyone who has walked down to a platform with a phone in hand has seen it: the blue dot freezes, starts jumping around, or disappears. A minor annoyance when you are looking for an exit. A serious problem when knowing where your security team is depends on it.

This article explains exactly what happens, why the problem is worse than it looks, and what the alternatives are.

A stopwatch twenty thousand kilometres away

GPS — and its equivalents Galileo, GLONASS and BeiDou — is not really a positioning system. It is a timing system. Each satellite continuously broadcasts a signal that says, in essence, “I am satellite number X and I sent this at exactly this instant.” The receiver compares that instant with its own clock, works out how long the signal took to arrive, multiplies by the speed of light and gets a distance. With four satellites in view, it solves a system of equations and knows where it is.

All of it depends on one condition: seeing the satellites. And there lies the problem, because those signals reach the Earth’s surface at a ludicrously low power, in the order of −130 dBm. They have travelled twenty thousand kilometres and arrive weaker than the receiver’s own thermal noise floor. The trick that makes GPS work is that the signal is coded so it can be recovered by correlation even while buried in noise.

That trick has a limit. A reinforced concrete ceiling attenuates the signal by 20 to 30 dB. Two floors underground, with steel beams and a station structure in between, attenuate it so much that there is nothing left to correlate.

What the phone does once the satellites are gone

Here is the part that does the most damage in an operational setting, and that rarely gets explained.

The phone does not say “I don’t know where I am.” It keeps returning a position, because the operating system blends several sources and falls back on whatever is left:

The last known position, which may be ten minutes old and from the station entrance.
The cellular network, triangulating by cell. In dense urban environments the error can run to hundreds of metres.
Databases of visible Wi-Fi networks, which place the device according to which access points it can see. Underground those are usually the station’s own access points, which are either wrong or absent in those databases.

The result is a plausible position, presented exactly like a good one. And that is the real problem: a control room looking at a dot on a map has no way of knowing whether that dot is worth anything.

If a system tells you your officer is in the concourse when they are actually on the opposite platform, you have made a decision on false data — and with more confidence than if you had had no data at all.

The alternatives, sorted by what they cost

When satellites cannot reach, you have to measure against something that is down there. The options come down to four families, and what separates them is less accuracy than the infrastructure they demand.

UWB anchors

Ultra-wideband is the accuracy benchmark: 10 to 30 centimetres under reasonable conditions. It works by measuring time of flight with very short pulses, which makes it fairly immune to multipath.

The price is infrastructure. It needs cabled, powered anchors distributed across every area you want to cover, plus an active tag on every person or asset to be located. In a warehouse or a factory, that is acceptable. In a metro network in service it means a construction project, permits, night-time work windows and months of calendar.

Bluetooth angle of arrival

BLE with AoA lowers the per-point cost compared with UWB and offers sub-metre accuracy. It still needs locators installed per zone, and it still needs a tag or emitting device on the person.

Wi-Fi by signal strength (RSSI)

This is the option that requires no new hardware, which is why people have been attempting it for twenty years. The idea is to estimate distance to each access point from how strongly its signal arrives.

The trouble is that received power depends on far too many things that are not distance: the phone model, antenna orientation, whether the device is in a hand or a pocket, how crowded the space is, and what obstacles sit in between. Two different phones at the same spot report very different readings. Compensating for that requires fingerprinting: walking the entire site taking samples and building a radio map. Which has to be rebuilt every time an access point moves or the furniture changes.

Typical errors run to tens of metres. In a station, that does not even tell you which level you are on.

Wi-Fi by time of flight (RTT/FTM)

This one changes the framing. Since 2016, the IEEE 802.11mc standard has included the FTM protocol (Fine Timing Measurement), which lets a terminal and an access point exchange a burst of frames and measure the round trip with nanosecond precision.

Since the signal travels at the speed of light, that time is a distance. And unlike RSSI, time does not depend on how you hold the phone.

The part that matters to a transport operator: nothing has to be installed. The access points are already mounted across stations — deployed to serve passengers or for internal use — and many of them only need the feature switched on. And the device doing the measuring is the corporate phone the officer already carries. Android has exposed this capability natively since Android 9.

Why the realistic figure is 5-10 metres, not 1-2

The academic literature reports 1 to 2 metres of error for Wi-Fi RTT. That figure is real, but it corresponds to direct line-of-sight conditions: terminal and access point see each other with nothing in between.

A platform is not that. There are columns, metal signage, a stationary train taking up half the station, and parallel walls throwing back echoes. When the signal arrives reflected rather than direct — a condition called NLOS, non-line-of-sight — it has travelled further than the straight-line distance, so the time measurement comes out longer and the system believes you are further away than you are.

And the error is not random: it is biased. Always in the same direction, never cancelling out across measurements.

That is why a serious system for this environment cannot stop at the raw measurement. It needs three more layers:

Detect the bad measurements. A reflected signal has an identifiable profile. Those measurements get flagged and lose weight rather than entering the calculation as if they were sound.
Filter with memory. A probabilistic filter carries the previous state forward, so an isolated spike does not teleport anyone thirty metres.
Use the layout. And this is the most powerful of the three.

The station layout is free information

A generic location system treats space as a continuous surface where the user could be anywhere. A metro station is nothing like that: it is a narrow, heavily constrained graph. Platform, concourse, interchange corridor, stairs, tunnel. People do not walk through walls or along the track.

Feeding that constraint into the estimator — what navigation calls map matching — has an effect that surprises people the first time they see it: the operational error ends up smaller than the geometric error of the individual measurements. Even if a given measurement is off by twelve metres, if those twelve metres point into a wall the system knows the position cannot be there, and corrects towards the most likely walkable space.

With all three layers applied, 5 to 10 metres in station is a figure that survives a measurement campaign. And in platform topology, five metres separate platform from concourse and ten separate one direction of travel from the other.

So is 5-10 metres any use?

It depends what for. To guide a robot around a warehouse, no. To coordinate an incident in a station, it is more than enough.

Nobody in a control room needs to know which tile their officer is standing on. They need to know where to send support, which entrance to use, and which of the available teams gets there first. For that, the difference between knowing the zone and knowing nothing is enormous; the difference between seven metres and seventy centimetres is irrelevant.

And there is a second half of the problem that fine accuracy does not solve either: reconstructing afterwards what happened. Who arrived first, how long it took, which way they came in. Without a record, that is one person’s word against another’s.