|
|
Journal of Global Positioning Systems Vol. 22, No. 2, 2026 |
Cover |
|
Journal of Global Positioning Systems
|
JGPS Team Structure, Copyright and Table of Contents |
|
JGPS Team Structure, Copyright
|
1. ED-SA-BiConvGRU: A Deep Learning Model for Ionosphere Prediction considering RMSE during Geomagnetic Storm in Australia |
|
Tu Yuan,Yize Zhang,Hanwen Yan,Junping Chen
Located at the core of the Equatorial Ionization Anomaly (EIA), the ionospheric Total Electron Content (TEC) in the Australian region experiences highly complex spatiotemporal evolution, which significantly degrades the accuracy of regional Global Navigation Satellite System (GNSS) positioning and navigation. Existing prediction methods—including traditional physical models and basic machine learning techniques—often overlook the intrinsic uncertainty and error dynamics of ionospheric products, failing to capture the abrupt spatiotemporal variations during severe geomagnetic storms
|
2. Mamba Augmented Adaptive Kalman Filter for Smartphone GNSS Navigation |
|
Divya Bhavsar,Hongzhou Yang
Smartphone GNSS measurements exhibit time-varying errors caused by multipath, non-line-of-sight reception, changing satellite availability, and device-dependent measurement characteristics. This study investigates a Mamba-augmented adaptive Kalman filter for group-level measurement covariance scaling in post-processed smartphone GNSS positioning. Using a centered 14-epoch window containing 20 measurement and contextual features, the Mamba selective state-space model predicts six variance scale factors for the GPS, GLONASS, and Galileo code and carrier-phase observation groups. These factors modify the variances produced by the existing -weighted RTKLIB stochastic model, while the process noise covariance and remaining positioning configuration are unchanged. The model is trained using ground-truth-referenced residuals from the Google Smartphone Decimeter Challenge (GSDC) 2022 training data, whereas test inference uses only features generated by the standard RTKLIB processing pipeline. The final configuration achieved a GSDC 2022 private-leaderboard score of 1.533 m, which would retrospectively correspond to third position in the published competition standings. Under a matched experimental configuration, Transformer and LSTM backbones obtained scores of 1.608 m and 1.626 m, respectively. Mamba also reduced trajectory-level RMS horizontal error for 59 of the 109 evaluated training drive-phone trajectories, although the benefit varied across trajectories and error metrics. The result was achieved using a GNSS-only differential post-processing framework without IMU integration, sensor fusion, or additional hardware. These findings demonstrate the potential of selective state-space modelling as a data-driven approach to temporal covariance adaptation in post-processed smartphone GNSS positioning.
|
3. Graph Convolutional Neural Networks for GNSS Multipath Mitigation in Urban Canyons via Observation and Coordinate Domains |
|
Chenkang Cui,Yan Xia,Bin Wang
Global Navigation Satellite System (GNSS) observations from simultaneously visible satellites in urban canyons are affected by a shared local environment and exhibit structured inter satellite relationships. This study adopts a Graph Convolutional Neural Network (GCNN) to model double differenced (DD) multipath labels and evaluates two correction paths in the observation and coordinate domains. A Global Positioning System (GPS)/BeiDou Navigation Satellite System (BDS) short-baseline pseudorange DD model is used to construct the labels. The GCNN employs a hybrid adjacency rule based on constellation membership, pseudorange-residual similarity, and elevation-angle similarity. By fusing features such as pseudorange residuals, carrier-to-noise-density ratio (C/N0), line-of-sight unit vectors and elevation angles, the model predicts multipath corrections for multiple satellites within the same epoch. The experiment uses the vehicle-borne Kowloon Tong (KLT) dataset collected by the Li-Ta Hsu group in the Kowloon Tong urban canyon. In the observation domain, the GCNN achieves a multipath prediction root mean square error (RMSE) of 3.15 m, compared with 4.85 m for the Convolutional Neural Network (CNN). After applying the GCNN predictions to DD pseudorange correction, the 3D positioning RMSE is reduced from 11.01 m for the DD baseline to 8.76 m. In the coordinate domain, the GCNN achieves a 3D correction RMSE of 7.85 m, lower than the 8.84 m obtained by CNN. These results show that the GCNN outperforms the CNN in both tasks. In the present test, the coordinate-domain path yields the smaller final positioning error, whereas the observation-domain path retains the advantage of introducing physically interpretable corrections at the measurement level.Keywords: Differential Global Navigation Satellite System (DGNSS); urban canyon; multipath mitigation; Graph Convolutional Neural Network
|
| To Be Completed |
Back Cover |
|
Journal of Global Positioning Systems
|