Assessing multi-GNSS PPP-AR using IGS all-frequency products and machine-learning-driven attribution analysis

Abstract Recent progress in the estimation of all-frequency bias corrections has substantially enhanced Precise Point Positioning with Ambiguity Resolution (PPP-AR), with the potential to achieve instantaneous convergence to centimetre-level accuracy. To investigate the performance of GNSS PPP-AR using the latest all-frequency products from the International GNSS Service (IGS) and to identify the observational factors associated with its performance, this study proposes a machine-learning-driven attribution analysis framework to evaluate PPP-AR performance and quantify the impact of individual observational factors. GPS+Galileo+BDS-3 observations from 440 IGS stations over a 28-day period in 2025 are processed using the uncombined Decoupled Clock Model (DCM). Positioning performance is evaluated in terms of convergence time and converged positioning accuracy. The framework incorporates Random Forest and XGBoost models to identify the observational factors that are most informative for PPP-AR performance. Results show that 98% of solutions achieve instantaneous convergence to a horizontal accuracy of 2.5 cm, while the average converged RMS error reaches 0.8 cm in the horizontal component and 2.0 cm in the vertical component when using the all-frequency products from Wuhan University. Performance discrepancies are observed among different GNSS products, receiver types, and antenna types, which are primarily attributed to differences in Galileo signal availability, bias correction coverage, signal-to-noise ratio levels, and multipath effects. The skyplot entropy is the most informative factor for horizontal positioning accuracy, whereas PDOP and mean satellite elevation are the factors most informative for vertical accuracy. These findings are validated through five-fold cross-validation, demonstrating good model stability and highlighting the potential of data-driven approaches to complement traditional GNSS analysis. In particular, the proposed framework reveals the relationship between observational factors and PPP-AR performance, thereby providing insights for optimizing processing strategies or configuring observation networks to improve PPP-AR positioning performance.