PPP-AR has long promised high accuracy from a single receiver, but early implementations depended on dual-frequency combinations, specific constellations and high-quality precise products. Multi-frequency and all-frequency bias corrections have widened ambiguity fixing to more signals from GPS, Galileo and BDS-3, yet performance still varies with product coverage, receiver type, antenna type, signal availability, signal-to-noise ratio (SNR) and multipath (MP). Conventional model-driven analysis can describe these effects, but it struggles to rank many correlated, non-linear factors in large networks. Based on these challenges, in-depth research is needed on which observational features are most informative for all-frequency multi-GNSS PPP-AR performance and how robust those associations remain across products and station configurations.
Researchers at York University and Sun Yat-sen University report (DOI: 10.1186/s43020-026-00223-x) the assessment in Satellite Navigation on 14 October 2026. They processed 440 IGS Multi-GNSS Experiment (MGEX) stations over 28 days in 2025, using two 6-hour sessions per day, and compared all-frequency products from Wuhan University (WUM) and the Centre National d’Études Spatiales (CNES; GBM0MGXRAP, or GBM). The study evaluates convergence time, converged accuracy and machine-learning-derived feature importance for PPP-AR solutions using two ensemble models and five-fold cross-validation. The analysis focuses on GPS+Galileo+BDS-3 observations and an uncombined DCM, and compares public all-frequency products and their impact on ambiguity fixing.
Using WUM products, 98% of solutions reached the 2.5 cm horizontal threshold at the first epoch, versus 90% with GBM products. Average converged root mean square (RMS) error was 0.8 cm horizontally and 2.0 cm vertically for WUM, and 1.2 cm and 2.6 cm for GBM; 95th-percentile errors were 1.6 cm/4.2 cm and 2.2 cm/4.9 cm, respectively. The gap is linked mainly to product completeness: WUM supported more ambiguity-fixable frequencies, while missing code and phase bias corrections reduced usable BDS-3 and Galileo observations for GBM. The attribution framework passed a known-answer test by distinguishing WUM from GBM solutions with an area under the curve (AUC) of 0.98 and 92% accuracy; the mean number of BDS-3 observations per satellite was the dominant separating feature. Receiver-type classification reached 95% (RF) and 96% (XGB), with SNR and MP standard deviation (STD) leading; antenna-type classification reached 92% and 93%, with SNR features dominant. For converged accuracy, skyplot entropy was the most informative feature for horizontal RMS error, whereas position dilution of precision (PDOP) and mean satellite elevation were most informative for vertical RMS error. Five-fold cross-validation supported model stability.
The authors said the results should be read as data-driven association, not proof of cause and effect. They said the agreement between RF and XGBoost, and across five folds, gives them confidence that skyplot entropy is a robust indicator for horizontal accuracy, while PDOP and mean elevation are the clearest indicators for vertical accuracy. They added that the framework is useful because it shows where product coverage, signal quality and station geometry leave detectable signatures after ambiguity resolution has already delivered centimetre-level solutions. They said this makes the ranking useful for diagnosis rather than as a substitute for physical modelling.
The findings have practical implications for PPP-AR users, product providers and network operators. More complete bias corrections and broader signal availability can raise the share of observations that participate in ambiguity resolution, while attention to skyplot geometry, PDOP and elevation can help flag sessions likely to show weaker vertical performance. The authors suggest the framework can support processing-strategy tuning and observation-network design, and can be extended as new signals and products arrive. It may also help users interpret differences among receivers and antennas without treating them as unexplained hardware effects. Because the method ranks associations rather than causes, follow-up physical modelling remains important for real-time and post-processed applications.
###
References
DOI
10.1186/s43020-026-00223-x
Original Source URL
https://doi.org/10.1186/s43020-026-00223-x
About Satellite Navigation
Satellite Navigation (ISSN: 2662-1363; ISSN: 2662-9291) Satellite Navigation is the official journal of the Aerospace Information Research Institute. The journal aims to report innovative ideas, new results, and progress in the theories, techniques, and applications of satellite navigation. The journal welcomes original articles, reviews and commentaries.