A nearby correction is useful only if it represents the air and signals at your receiver.
You move a survey receiver from one end of a site to the other. Both observations display a fixed solution, but the second disagrees with a check mark. Before blaming the receiver, ask a more useful question: which part of the correction was actually measured nearby, and which part was inferred for this location?
Network real-time kinematic positioning, usually shortened to network RTK, uses observations from multiple known reference stations to help a rover resolve carrier-phase ambiguities and estimate its position. The network can represent the spatial changes in atmospheric and other range errors better than a single distant reference under suitable conditions. It does not directly measure the path between every satellite and every rover.
This guide follows the inference from the reference stations to the rover. Start with a two-station calculation, test it against a station left out on purpose, and see what changes at an edge, during a local atmospheric disturbance or when one reference station stops reporting. The numerical examples are deliberately simple; they explain the checks, not a service’s particular algorithm.
Geodesy subject guide · Learning that transfers into life · How X Works hub
Choose your reading route
Route 1 · Chapters 1–2: Understand the observation
Start with one reference and one rover
Trace what differencing can cancel and what remains spatially variable.
Route 2 · Chapters 3–5: Build and stress-test an interpolation
Interpolate two values before adding complexity
Calculate weights, inspect coverage and separate atmospheric components.
Route 3 · Chapters 6–8: Follow the delivered correction
Separate the delivery format from the inference
Check service format, time tags and correlated errors.
Route 4 · Chapters 9–10: Diagnose a discrepancy
Use a holdout that tests the question you are asking
Use holdouts and independent marks to distinguish plausible causes.
Route 5 · Chapters 11–12: Transfer and decide
A station outage can change the supported region
Work a changing-network case and state what the evidence supports.
Full chapter index · Independent worked case · Questions and practical checks
Chapter index
Understand the observation
Start with one reference and one rover
The network estimates differences, not a weather map
Build and stress-test an interpolation
Interpolate two values before adding complexity
See the difference between being near and being supported
A line can pass through every reference and still miss the rover
Follow the delivered correction
Separate the delivery format from the inference
Check the clock on the correction, not just the connection
Do not count shared station errors twice
CHAPTER 1 OF 12 · Understand the observation
1. Start with one reference and one rover
Picture two receivers tracking the same satellite at nearly the same time. Both observe a carrier whose phase is sensitive to distance, clocks, atmosphere, instrument effects and an unknown integer number of cycles on each continuous track. Because the receivers see the same satellite, forming a difference between the observations removes some common terms under a consistent model. A second difference across satellites removes another common clock term in the usual idealised model. Some terms survive because the receivers do not see identical signal paths.
In a short-baseline thought experiment, suppose one satellite’s effective atmospheric contribution is +48 mm at the base and +51 mm at the rover. Differencing leaves +3 mm for that component. If a base farther away has +22 mm while the rover still sees +51 mm, the difference leaves +29 mm. These are example contributions to one observation, not position errors. The geometry and the other measurements determine how they affect the solution.
The known base coordinate is an important part of the construction. If it is wrong, a consistent-looking baseline can inherit a coordinate problem. Likewise, a fix is a decision about a set of integer unknowns under a model. It does not check the base frame, the rover antenna height, or whether an atmospheric assumption held at the rover.
Keep one sketch: a satellite at the top, two slant paths through the atmosphere and two ground receivers. Label the terms that are shared and those that change with the path. When you add reference stations, you are filling in a picture of changing terms across the network. You are not physically shortening the rover’s signal path through the atmosphere.
CHAPTER 2 OF 12 · Understand the observation
2. The network estimates differences, not a weather map
Reference stations have coordinates supplied in a declared reference frame. They observe multiple satellites and signals. The processing uses those observations and models to estimate or represent spatially correlated range errors once the relevant network ambiguities and other unknowns are suitably treated. A rover can then use information adapted to its neighbourhood rather than assuming one far-away base has the same residuals.
A useful teaching convention is to call a path residual positive when it makes the modelled observation too short in an observed-minus-computed comparison. If the network estimates +18 mm for a particular observation at the rover, a simplified corrected residual subtracts +18 mm. A product may publish the opposite sign as a correction to add. Those can describe the same action, but mixing them leaves twice the intended error in the simple example.
Real services do not necessarily expose a single scalar atmospheric residual. They can distribute observations or corrections with satellite-, signal- and epoch-specific definitions. Some terms are dispersive, which means they change with signal frequency; other terms are not treated that way at the precision of the model. Read the interface and product description before trying to replicate the example as an implementation.
A good mental model is a prediction with support, not a guarantee. The network can observe what happens at its stations. At the rover it relies on a spatial and temporal model. Even an excellent fit at the stations may miss a local feature between them. The aim of the checks below is to find the conditions under which that matters.
CHAPTER 3 OF 12 · Build and stress-test an interpolation
3. Interpolate two values before adding complexity
Imagine two references along a straight 40 km corridor. At the western end the comparable residual is +12 mm; at the eastern end it is +32 mm. At a rover 10 km from the west, a linear model puts it one quarter of the way across. Use weights 0.75 and 0.25: the prediction is 0.75 × 12 + 0.25 × 32 = +17 mm.
Check the endpoints. At the west the weights become one and zero, reproducing +12 mm. At the east they become zero and one, reproducing +32 mm. Check the midpoint: the predicted +22 mm lies halfway between them. These are basic implementation tests. They tell you the weighting behaves as intended, not that the atmosphere actually follows a straight line.
Suppose a temporary independent reference at the rover location measures +24 mm for the same observable and epoch. The leave-out discrepancy is +7 mm: observed residual minus predicted residual. It is evidence about the interpolation at this place and time. It does not identify whether the discrepancy came from curvature in the field, an instrument effect at the temporary station or a difference in processing.
Transfer the method to a new pair: −8 mm at the west, +16 mm at the east, and a rover 30 km from the west. Its expected value is 0.25 × (−8) + 0.75 × 16 = +10 mm. The sign of the predicted value is not a vote taken among stations; it follows the weights and the signed terms.
CHAPTER 4 OF 12 · Build and stress-test an interpolation
4. See the difference between being near and being supported
Now put the rover 10 km east of the eastern reference. The same line gives a fractional distance of 1.25, weights −0.25 and +1.25, and a prediction of +37 mm from the original +12 and +32 mm endpoints. This is extrapolation. A plausible number still comes out, but the prediction has no station beyond the rover to constrain how the field behaves there.
In two dimensions, three suitably separated stations can define a plane in the simplest local illustration. A point inside their triangle can have nonnegative weights that sum to one. At an edge or beyond it, continuing the plane can require negative weights. That does not make extrapolation automatically unusable, but the limits, assumptions and uncertainty deserve separate treatment.
Geometry matters beyond a yes-or-no boundary. A skinny triangle made from near-collinear references may poorly constrain change across the corridor even if a map shows three stations. If a coastal site has all its usable references on the landward side, the closest station might still leave the seaward gradient weakly observed. Look at the participating stations, not the provider’s total station count.
Ask which direction the evidence actually spans. In a long narrow network, shifting the rover across the corridor may be riskier than moving the same distance along it. A useful qualification states the geometry, active station set and boundary test. It does not turn a nominal coverage polygon into a uniform accuracy statement.
CHAPTER 5 OF 12 · Build and stress-test an interpolation
5. A line can pass through every reference and still miss the rover
Return to the 40 km corridor. Suppose both endpoints measure +20 mm. A line predicts +20 mm everywhere. If a local disturbance raises the comparable residual to +35 mm near the centre, neither endpoint reveals it. The zero residual at the stations is a property of fitting the chosen model at those stations. It is not evidence that the midpoint had zero prediction error.
The atmospheric terms also behave differently. Ionospheric effects are frequency-dependent, while the neutral-atmosphere delay is modelled differently and depends on the path through changing pressure, temperature and water vapour. An interpolation model that works well for one residual component or elevation range may not work for another. An all-in-one score can hide which component failed.
Elevation and topography change the problem. A ridge reference and a valley rover may sit at similar map distances from their neighbours yet have different atmospheric paths. A low-elevation satellite takes a different slant path from a high-elevation one. Changing the elevation mask may reduce difficult observations but also changes positioning geometry; a cleaner plot is not automatically a stronger position estimate.
Do not assume more stations always fix the difficulty. A third station in the centre could expose the bump, but an extra reference beside one endpoint may add little information about it. What helps is a measurement placed to distinguish competing spatial explanations, processed in a way that lets you compare like with like.
CHAPTER 6 OF 12 · Follow the delivered correction
6. Separate the delivery format from the inference
Network correction services can distribute information in different forms. One familiar design provides observations associated with a virtual reference station near the rover. Other designs transmit observations and network information for a reference and auxiliary stations, or correction parameters from which a compatible rover constructs its model. The messages and processing assumptions matter more than the resemblance of their labels.
A virtual reference is not a receiver that physically made observations next to the rover. It represents observations constructed using real network data and a model. The short numerical distance to a virtual point is therefore not the same as the distance to independent field evidence. A report should distinguish the two if the user might treat a virtual baseline length as a local measurement guarantee.
Suppose a rover moves several kilometres but the stream continues to use a virtual position established earlier. Whether this is supported, when the stream should be refreshed and what happens to ambiguities depend on the service and rover behaviour. Avoid a universal rule about the allowed distance. Record the stream type, reference coordinate and any documented transitions.
Treat stream switches as changes to the experiment. A switch may alter the participating station set, the reference convention or the observable definitions. Compare overlapping observations and independent check points where possible. “The correction stream stayed connected” does not show that the model and reference stayed equivalent.
CHAPTER 7 OF 12 · Follow the delivered correction
7. Check the clock on the correction, not just the connection
Let a simplified residual at a location change steadily by +4 mm every second for a short interval. If the model applies an estimate from three seconds earlier without a suitable prediction step, the example leaves +12 mm in the observation. That is a teaching calculation, not an operational correction-age limit. Real changes vary by signal, environment, service and time.
Distinguish the epoch a value describes from the time it arrived. A packet that arrives now may describe an earlier state. A stream can also send values with different effective update times. Write both times in the log and retain receiver quality flags. Without the distinction, an apparently spatial problem can be a time mismatch.
A dropout and a recovery are separate events. A network station may disappear while the rover keeps tracking, or the rover may lose observations while corrections continue arriving. In the first case the spatial model can change even though the rover arcs survive. In the second, the rover may have to establish new ambiguities even if the reference network did not change.
Look for steps when coverage changes. Before and after a handover, repeat a known position or test an independent observation, using comparable frame and antenna information. A precise-looking result immediately after reconnection is not by itself evidence that the ambiguity state, correction age and spatial model are all valid.
CHAPTER 8 OF 12 · Follow the delivered correction
8. Do not count shared station errors twice
Two network estimates can look like independent checks and share most of their inputs. If both streams use station A, the same satellite products and similar atmosphere assumptions, their agreement is weaker evidence against errors in those shared components. Ask which input was actually changed and what error you intended to test.
For a toy uncertainty calculation, use the weights 0.75 and 0.25 from the corridor case. Each station residual has a standard deviation of 8 mm. If the errors are independent, the prediction variance is 0.75² × 64 + 0.25² × 64 = 40 mm², giving about 6.32 mm. If the errors have covariance +48 mm², add 2 × 0.75 × 0.25 × 48 = 18 mm², raising the standard deviation to about 7.62 mm.
Neither result includes a separate allowance for model mismatch at the rover. A perfectly known value at each endpoint still does not reveal an unsampled bump. If you have a defensible spatial prediction uncertainty, account for it without assuming independence where it is not justified. Keep the numerical and structural uncertainties visible as different problems.
Now use the extrapolation weights −0.25 and +1.25. With independent errors of the same size, the variance rises to 104 mm². This is not a network performance formula. It shows why extrapolation can amplify some observation errors and why covariance terms can have a different effect when a weight changes sign.
CHAPTER 9 OF 12 · Diagnose a discrepancy
9. Use a holdout that tests the question you are asking
To test spatial prediction, leave out a station completely from the fit for the relevant test interval. Compare the prediction at that station with a reference quantity derived compatibly there. A station that contributed to the solution is useful for model diagnostics, but it is not an independent spatial holdout.
Design the test with more than one easy interior case. Compare centre, edge and sparse directions; quiet and variable intervals; and configurations where a station goes offline. Label the weather or ionospheric proxy if you use one, but avoid substituting it for the error measurement. The point is to see whether the model’s predictions generalise under the actual changes of interest.
Look at signed residuals, not only a single root-mean-square statistic. A mean error near zero can hide positive differences during one part of the day and negative ones during another. An acceptable overall score can hide a poor corner of the network. Group by geometry, signal, time and active station set where the data support that resolution.
A good test should be able to fail. Write down which discrepancy would prompt an investigation, what independent measurements would discriminate among explanations and how exclusions are documented. Keep both failed and successful cases. If you only keep intervals where the rover fixed, the result will not describe the full operating conditions.
CHAPTER 10 OF 12 · Diagnose a discrepancy
10. When a fixed rover misses a known mark
Begin with the comparison rather than the diagnosis. Confirm the coordinate reference, epoch where applicable, antenna height and reference point, and correct identification of the check mark. Confirm the independent reference is fit for the accuracy you are trying to demonstrate. A disagreement in the vertical may come from height conventions before any interpolation issue enters the analysis.
Then look at the timing and geometry. Were the questionable readings acquired after a stream change, with fewer references, near an edge or under poor sky visibility? Compare raw observations or receiver diagnostics if available. Local multipath, signal interruption or a wrong integer fix can explain a discrepancy without requiring a network-wide atmospheric failure.
For a changed example, suppose two nearby rovers at known marks both develop a similar signed offset after a documented network switch, while repeated occupations before the switch agree. That makes a common reference, correction or processing change worth testing. It still does not isolate atmospheric interpolation, because both rovers may share the same reference change or processing path.
If one rover alone shows the error, compare antenna setup, nearby surfaces and arc continuity before attributing it to the service. If independent observations are missing, state that the reported coordinates disagree with the check under the recorded conditions. Leave the specific cause open rather than choosing the most technical-sounding explanation.
CHAPTER 11 OF 12 · Transfer and decide
11. A station outage can change the supported region
Consider references at 0, 20 and 40 km and a rover at 35 km. At one epoch the residuals are +10, +17 and +30 mm. Using the two bracketing stations, the fraction from 20 toward 40 km is 0.75. The prediction is 0.25 × 17 + 0.75 × 30 = +26.75 mm.
If the eastern station drops out, the surviving references at 0 and 20 km can be extended to 35 km only by extrapolation in this model. The slope is (17 − 10)/20 = 0.35 mm per kilometre. The prediction becomes 10 + 35 × 0.35 = +22.25 mm. The +4.50 mm difference between the two predictions is a consequence of changed support, not proof that the atmosphere itself jumped.
Suppose an independent check then gives +28 mm. The interpolation error would be +1.25 mm and the extrapolation error +5.75 mm for this fictional observable. If the check were +21 mm instead, the extrapolated value would be closer. It is the design and the held-out evidence, not the word interpolation, that justify a claim about quality.
A practical alert can therefore distinguish “network degraded or geometry changed” from “position error confirmed.” The first tells you to revisit the evidence; the second requires a check. Preserve the active station set and exact time of the transition so you can reproduce the comparison later.
CHAPTER 12 OF 12 · Transfer and decide
12. Transfer the idea to measurements off the survey site
Imagine estimating temperature in a greenhouse from sensors at its ends. A midpoint estimate may be good until a heater or open door creates a local pattern. A third sensor near an end cannot resolve every new feature. Add a sensor where the two plausible temperature fields would predict different readings. That is the same logic as positioning a useful holdout reference.
Now return to a rover and ask the same three questions. What variables did the network actually observe? Which relationship did it assume between stations? What independent information could reveal a violation? If the only answer to the last question is “another result from the same service,” inspect the shared inputs before treating the agreement as independent.
A concise result is more useful than a blanket assurance: “Under these active reference stations and time tags, three interior check occupations agreed with the independent mark within the stated limit. Edge and post-dropout performance were not established.” The next reader knows what was checked and which extension still needs evidence.
Take the habit away: map the support before trusting the prediction. An interpolated correction can substantially improve the use of nearby observations, but the correction is still an estimate for a particular place, epoch and model. The right question is not whether it looks smooth. It is whether independent evidence supports carrying it to the location where a decision will be made.
Independent worked case and evidence checks
An independent fictional test uses references A at 0 km and B at 30 km and reserves C at 12 km as a check. On a shared observable the reference residuals are +9 and +24 mm. The interpolation fraction is 0.4, so the predicted C value is 0.6 × 9 + 0.4 × 24 = +15 mm. C independently reads +19 mm, leaving a signed prediction discrepancy of +4 mm.
Repeat with a different observation window: A and B both read +9 mm and C reads +18 mm. The prediction becomes +9 mm and the discrepancy grows to +9 mm despite the endpoints agreeing. The endpoint agreement cannot detect a feature between them. Before attributing the new discrepancy to atmosphere, check the C instrument and matched time and signal definitions.
Now place a rover at 36 km. With the original endpoints and the same linear model the extrapolation predicts +27 mm. A new reference at 42 km reads +21 mm. Interpolating between 30 and 42 km predicts +22.5 mm instead. The two estimates disagree because the extra station constrains a change beyond the original network, not because one arithmetic step was faulty.
For an actual acceptance decision, do not use the difference between these models as the only uncertainty. Record whether the newer reference is compatible, whether the rover observation is time-matched, and whether a held-out mark supports the model under comparable conditions. If that last check has not been performed, report the available evidence without converting the estimated correction into a demonstrated positional accuracy.
Worked comparison
| Situation | First check | Unresolved question |
|---|---|---|
| Inside a wide triangle | Use matched spatial holdouts | Is there a local feature between stations? |
| Near or beyond the edge | Check active stations and extrapolation | What constrains the missing direction? |
| After a dropout | Inspect station set and correction epochs | Did the spatial model or rover state change? |
| Two streams agree | List shared inputs | Which errors could both share? |
Questions and practical checks
Does network RTK measure the atmospheric error at my receiver?
The receiver makes its own observations. The network estimates information for its location using reference measurements and a model; it does not directly observe every rover’s atmospheric paths.
Is a virtual reference a real nearby receiver?
No. It is a reference constructed from network information. Its coordinate is not proof of independent observations at that spot.
Does a fixed solution prove the correction was correct?
No. It indicates an ambiguity decision under the processing model. Independently check the position and the conditions that support the model.
Are service-edge results always unreliable?
No. They require relevant evidence. Their geometry and assumptions may differ from those of a well-supported interior location.
What should I keep in a check log?
Record station participation if available, stream type, correction epoch and arrival time, reference frame, antenna setup, rover status and the independent comparison.
Continue with the connected subjects
Understand reference and ambiguity choices
GNSS double differences: what cancels and why errors are correlated
Integer ambiguity search: separate a chosen candidate from a validated fix
PPP outages: distinguish tracking loss from correction interruption
Check atmosphere and local geometry
GNSS zenith delay: compare elevation mapping and height trade-offs
GNSS tropospheric gradients: understand direction and coverage
GNSS multipath: check stable local errors before blaming a network
Coordinate epochs: align the reference before comparing positions
Sources and scope
Original comparison: using network RTK corrections in differential GNSS positioning
Institute of Navigation: spatial residual modelling for network RTK corrections
The numerical cases in this guide are original fictional teaching examples, not measurements from a real site. This is an educational explanation, not a field safety instruction, engineering approval or an official certification. For a real project, use qualified practitioners, current product documentation and site-specific evidence. Sources were checked on 4 November 2026.