GNSS Research Identifies Factors Behind Centimetre-Level Positioning Accuracy
Two sets of satellite correction products produced noticeably different positioning results across an international monitoring network, despite both being used with advanced centimetre-level positioning techniques. In research covering 440 GNSS stations, one set of products enabled 98% of solutions to meet a 2.5 cm horizontal accuracy threshold at the first processing epoch, compared with 90% for the alternative.
The difference was largely associated with the availability of correction data for individual satellite signals. Researchers from York University and Sun Yat-sen University have used machine learning to investigate which other observational characteristics are most closely associated with positioning accuracy, with findings relevant to precision surveying, construction machine guidance and infrastructure monitoring.
Their study, published in Satellite Navigation, examined Precise Point Positioning with Ambiguity Resolution (PPP-AR), using observations from GPS, Galileo and China’s BeiDou-3 satellite constellation.
The researchers found that the distribution of satellites across the sky, their elevation and the geometric strength of the observations were associated with the accuracy achieved after positioning solutions had converged. The findings provide a more detailed picture of the conditions affecting precision satellite positioning, although the research was conducted using fixed monitoring stations rather than moving machinery.
Briefing
- Researchers analysed 440 international GNSS monitoring stations over 28 days in 2025.
- The study examined combined GPS, Galileo and BeiDou-3 observations using all-frequency PPP-AR processing.
- Wuhan University correction products achieved the 2.5 cm horizontal threshold at the first processing epoch in 98% of solutions, compared with 90% for the CNES products tested.
- Average converged horizontal positioning errors were 0.8 cm and 1.2 cm respectively, with vertical errors of 2.0 cm and 2.6 cm.
- Machine learning identified satellite distribution, signal quality and observation geometry as useful indicators of positioning performance.
Understanding Precise Point Positioning
Global Navigation Satellite Systems (GNSS) have become an established part of modern construction and surveying. They support everything from establishing site control points and mapping earthworks to guiding graders, excavators and other equipment fitted with satellite-based positioning systems.
Achieving centimetre-level positioning generally requires considerably more than the basic satellite navigation signals used by consumer devices. Errors associated with satellite clocks, orbital information, atmospheric conditions and signal measurements must be accounted for through additional processing and correction data.
PPP uses precise satellite orbit and clock information to improve positioning accuracy without requiring the same local reference-station arrangement as conventional Real-Time Kinematic (RTK) positioning. PPP-AR extends this approach by resolving the integer ambiguities associated with carrier-phase measurements, improving the speed and reliability with which precise solutions can be obtained.
The availability of additional frequencies and satellite constellations has expanded the number of observations that can contribute to this process. However, receiving more signals does not automatically guarantee a better result. The correction products must also contain the information needed to use those signals effectively.
The latest research examined that relationship using an uncombined processing model incorporating GPS, Galileo and BeiDou-3 observations. The analysis covered 28 days of measurements from 440 stations in the International GNSS Service Multi-GNSS Experiment network, with two six-hour processing sessions examined each day.
Correction Product Coverage and Positioning Accuracy
The researchers compared all-frequency correction products from Wuhan University (WUM) and France’s Centre National d’Études Spatiales (CNES), using the GBM product identified in the study.
The differences were measurable in both the initial positioning performance and the accuracy of the converged solutions.
With WUM products, average converged root mean square (RMS) positioning errors were 0.8 cm horizontally and 2.0 cm vertically. The corresponding results using GBM products were 1.2 cm horizontally and 2.6 cm vertically.
At the 95th percentile, horizontal and vertical errors reached 1.6 cm and 4.2 cm respectively for WUM, compared with 2.2 cm and 4.9 cm for GBM.
The first-epoch results were particularly striking, although they relate to the study’s specific processing approach and evaluation criteria rather than the time a receiver would require to establish an accurate position when first switched on in the field.
The researchers attributed much of the difference between the products to the completeness of their correction information. WUM supported ambiguity resolution across a larger number of signal frequencies, while missing code and phase bias corrections in the GBM products limited the usable observations from certain BeiDou-3 and Galileo signals.
A receiver may be capable of tracking numerous frequencies and constellations, yet the processing system cannot necessarily exploit every observation unless suitable correction information is available.
The results represent the performance of the particular products and configurations examined during the research period. They should not be interpreted as a permanent ranking of correction providers, whose products and signal coverage may evolve.
Using Machine Learning to Identify Positioning Errors
A second part of the research examined whether machine learning could identify the observational characteristics associated with stronger or weaker positioning results.
The researchers employed Random Forest and XGBoost, two ensemble learning methods capable of examining complex relationships among multiple variables. Their models considered characteristics including satellite geometry, signal-to-noise ratio, multipath effects, receiver and antenna types, and the availability of observations from different constellations.
The approach was tested against a classification problem with a known answer: determining whether positioning solutions had been produced using WUM or GBM correction products.
The framework achieved an area under the receiver operating characteristic curve of 0.98 and classification accuracy of 92%. The mean number of BeiDou-3 observations per satellite emerged as the strongest feature separating the two sets of solutions.
The models could also distinguish between receiver and antenna types with relatively high accuracy. Receiver classification reached 95% using Random Forest and 96% using XGBoost, while antenna classification achieved 92% and 93% respectively. Variations in signal-to-noise ratio and multipath measurements were particularly informative for receiver classification, while signal-to-noise characteristics dominated the antenna results.
These classifications do not establish that particular receivers or antennas caused better positioning outcomes. They demonstrate that different equipment configurations leave identifiable patterns in the satellite observations, allowing researchers to investigate performance differences across large datasets.
Satellite Geometry and Positioning Performance
The research identified different observational features associated with horizontal and vertical positioning accuracy.
For horizontal RMS error, the most informative feature was skyplot entropy, a measure describing characteristics of how satellite observations are distributed across the visible sky. The spatial arrangement of satellites affects the geometric strength of positioning measurements, particularly when observations are concentrated in limited areas rather than spread across different directions.
Vertical positioning accuracy showed a different relationship. Position Dilution of Precision (PDOP), which describes the geometric strength of the satellite configuration, and mean satellite elevation emerged as the most informative features in the machine-learning analysis.
Satellite elevation influences the path that signals travel through the atmosphere and their exposure to obstructions and reflections. Geometry also affects how measurement errors propagate into the estimated position.
Vertical positioning can be particularly sensitive to satellite geometry, making elevation estimates more difficult to maintain consistently than horizontal coordinates.
For construction surveying and machine guidance, vertical accuracy is often especially demanding. A grader working to a specified road formation level, for example, must maintain a reliable elevation reference as it moves across the site. Small positioning discrepancies can affect the quality of the finished profile, particularly where tight tolerances apply.
The study’s results suggest that examining observation geometry alongside conventional positioning quality indicators could help identify conditions associated with weaker performance. The researchers established statistical associations rather than demonstrating that changing any single feature would produce a predictable improvement.
Applications in Construction and Infrastructure
The research was conducted using observations from fixed GNSS monitoring stations, but several of the technical considerations are familiar in precision positioning applications across construction and infrastructure.
Machine control systems must often work in environments where satellite visibility is restricted by cuttings, buildings, bridges, stockpiles or nearby equipment. Reflected signals can also introduce multipath errors, particularly where large metallic surfaces or other obstructions affect reception.
Moving machinery introduces additional challenges associated with changing satellite visibility, receiver dynamics and the integration of GNSS with other positioning sensors.
The study does not quantify these effects in construction environments. It does, however, provide evidence that the completeness of correction products and the characteristics of available observations can influence the quality of PPP-AR solutions.
For surveying organisations and positioning service providers, the findings offer a basis for examining why different receivers, antennas or correction products may produce different results under apparently similar conditions.
They may also inform the design and management of permanent GNSS reference networks used for surveying, geospatial monitoring and positioning services.
In infrastructure monitoring, where repeated observations may be used to assess movement or deformation, distinguishing changes in measurement quality from actual changes in position is particularly important. Statistical assessment of satellite geometry and observation characteristics could provide an additional diagnostic tool, although its effectiveness for detecting structural movement was not evaluated in this study.
Improving the Reliability of Precision GNSS
The researchers tested the stability of their machine-learning analysis using five-fold cross-validation and compared results across the two modelling approaches. The agreement between methods supported the consistency of the feature rankings within the dataset.
The authors emphasised that these rankings describe associations rather than direct cause-and-effect relationships. Machine learning can identify which variables are most informative when explaining observed performance differences, but physical modelling remains necessary to understand the mechanisms involved.
The findings could help positioning service providers diagnose differences between correction products and identify observational conditions associated with weaker solutions. Further evaluation would be needed to establish how well the same relationships hold for real-time positioning, moving receivers and more obstructed operating environments.
The study also exposes a limitation of headline positioning specifications. Two processing configurations can both deliver centimetre-level results while differing in the availability of usable signals, the consistency of vertical measurements and the conditions under which accurate solutions can be maintained.
For surveying and infrastructure applications, understanding those differences is part of selecting and managing a reliable positioning service. The next challenge is establishing how consistently the same relationships hold when receivers move from permanent monitoring stations into the more demanding environments of working construction sites.
Key Industry Questions
- What is PPP-AR satellite positioning? Precise Point Positioning with Ambiguity Resolution uses precise satellite correction information and carrier-phase ambiguity fixing to achieve high-accuracy GNSS positioning without relying on a conventional local RTK base-station arrangement.
- How does PPP-AR differ from RTK positioning? RTK typically uses measurements from a local reference station or network to resolve positioning errors. PPP-AR relies on precise satellite products and ambiguity-resolution techniques, although the available correction services and processing methods vary.
- Can PPP-AR achieve centimetre-level accuracy? Yes. The study demonstrated average converged horizontal RMS errors as low as 0.8 cm under its tested conditions. Actual performance depends on the correction products, equipment, observations and operating environment.
- Why do satellite correction products affect positioning accuracy? Correction products supply information needed to account for satellite-related errors and resolve carrier-phase ambiguities. Missing bias corrections can prevent certain signals from contributing fully to the positioning solution.
- Why is vertical GNSS accuracy often weaker than horizontal accuracy? The geometry of visible satellites and the propagation of measurement errors generally make vertical estimates more difficult to constrain. Satellite elevation and dilution of precision are important considerations.
- What is multipath interference in GNSS? Multipath occurs when satellite signals reach a receiver after reflecting from surrounding surfaces. These additional signal paths can distort measurements and reduce positioning accuracy.
- Can machine learning improve GNSS positioning? Machine learning can identify patterns associated with positioning errors and help diagnose performance differences. This research evaluated feature importance rather than demonstrating a machine-learning correction system that directly improves positioning accuracy.
- Does the research prove better performance for autonomous construction machinery? No. The study used fixed monitoring-station observations. Validation on moving construction equipment would be necessary before drawing conclusions about dynamic machine guidance or autonomous operations.
Strategic Takeaways
- Correction product completeness affects usable satellite observations. Multi-frequency receiver capability delivers its full benefit only when the processing system has the required correction information.
- Horizontal and vertical positioning accuracy have different sensitivities. Satellite distribution was particularly informative for horizontal accuracy, while PDOP and elevation were stronger indicators for vertical performance.
- Hardware comparisons require observational context. Receiver and antenna characteristics can be associated with measurable signal-quality differences, but those associations do not establish causation.
- Machine learning offers a diagnostic approach. Feature ranking can help investigate positioning performance across large monitoring networks without replacing established GNSS physical models.
- Construction applications require separate validation. Fixed-station research provides useful technical evidence, but dynamic machinery, obstructions and multipath introduce operating conditions that were not directly tested.
