Answers are grounded in this paper and cite their sources.
Multipath error is one of the primary factors limiting high-precision Global Navigation Satellite System (GNSS) positioning, particularly in complex urban canyon environments. Conventional GNSS Multipath Hemispherical Map (MHM) methods rely on discrete spatial grids, which limit spatial resolution and lack the capability to jointly represent multi-system and multi-frequency observations within a unified framework. Deep learning-based methods can provide continuous fitting of multipath errors. However, standard neural networks exhibit spectral bias, which can limit their ability to represent high-frequency spatial variations caused by severe obstructions and strong reflections. To address these challenges, this study proposes an implicit neural representation framework, termed INR-MHM, for continuous modeling of multipath errors for multi-system and multi-frequency GNSS. The proposed method reformulates the discrete multipath field into a continuous coordinate-based function. Specifically, this model introduces a multi-scale Fourier spatial encoding technique to map low-dimensional spatial coordinates into a high-frequency feature space, thereby mitigating the spectral bias of neural networks. Simultaneously, a context-gated feature fusion mechanism is designed to dynamically integrate multidimensional attributes, such as spatial geometry, signal frequency, and signal quality. Finally, a lightweight nonlinear regression head maps the fused representation to the estimated Single-Differenced (SD) carrier-phase multipath correction. The models were evaluated over independent 21-day testing periods at two monitoring stations: HF23, a slow-deformation landslide monitoring station, and GZY12, a highly obstructed highway slope monitoring station. Compared with the uncorrected SD carrier-phase residuals, INR-MHM reduced SD carrier-phase residual Root-Mean-Square (RMS) by 34.7%–72.2% at HF23 and 52.4%–61.0% at GZY12 across multi-system and multi-frequency GNSS observations, involving Global Positioning System (GPS), BeiDou Navigation Satellite System (BDS), Galileo navigation satellite system (Galileo), and GLObal NAvigation Satellite System (GLONASS). In the position domain, the Three-Dimensional (3D) positioning RMS errors were reduced from 7.8 mm to 4.7 mm at HF23 and from 7.2 mm to 3.8 mm at GZY12. At the highly obstructed GZY12 station, INR-MHM also increased the ambiguity fixing rate from 89.7% to 99.7%. Comparisons with MHM, Trend-Surface MHM (TMHM), and XGBoost demonstrated that INR-MHM generally achieved lower residual RMS values and the lowest 3D positioning RMS at both stations, with more substantial improvements under the highly obstructed condition. In addition, the ablation study indicated that multi-scale Fourier encoding was the main component contributing to residual correction accuracy, while context-gated feature fusion provided an adaptive mechanism for integrating heterogeneous spatial, system-frequency, and signal-quality features. Overall, the experimental results suggest that INR-MHM can improve multipath modeling and mitigation performance across heterogeneous multi-system and multi-frequency GNSS observations at the two tested monitoring stations, with larger improvements under the highly obstructed condition.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: