Answers are grounded in this paper and cite their sources.
Smartphone-based GNSS positioning is highly susceptible to multipath interference, a major source of positioning errors of approximately 20–30 m in challenging environments. Unlike high-end GNSS receivers equipped with choke ring antennas and advanced correlators, smartphones lack specialised hardware for multipath mitigation, making software-based detection essential. This study proposes a physics-informed machine learning framework for multipath detection using raw GNSS measurements from the Google Smartphone Decimeter Challenge dataset, collected from multiple Android devices in real-world environments. Ground-truth labels were generated and validated using NovAtel SPAN RTK residual anchors, ensuring physically meaningful supervision without circular feature dependency. The resulting labels achieved a precision of 1.000, recall of 0.912, F1-score of 0.954, and an overall agreement of 92.2%. Five supervised classifiers—Random Forest (RF), Logistic Regression, Support Vector Machine, Gradient Boosting, and Naive Bayes—were evaluated under imbalanced (90:10), moderately balanced (60:40), and fully balanced (50:50) class distributions. RF achieved the best performance in the balanced setting, with an MCC of 0.645, ROC-AUC of 88.57%, and F1-score of 81.47%, outperforming a C/N₀-only baseline by 0.611 MCC points. SHAP analysis identified device encoding, carrier phase smoothness, and Doppler rate anomaly as the most influential features. Leave-one-device-out validation across 20 devices produced a median RF F1-score of approximately 83%, while a physics-only ablation improved generalization to unseen hardware. Integrating predicted multipath probabilities into soft weighted least-squares positioning reduced the median positioning error from 3.71 m to 2.50 m and increased solution availability from 66.2% to 99.2%, demonstrating the practical value of probabilistic multipath mitigation.
No takes yet. Share an insight, caveat, or question.
Agarwal et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: