High-precision Global Navigation Satellite System positioning relies heavily on millimeter-level carrier-phase tracking, yet multipath interference remains one of the most stubborn physical barriers limiting operational accuracy in challenging observation environments. To address the fundamental resolution constraints of conventional spatial correction grids, researchers have introduced an implicit neural representation framework termed INR-MHM. This data-driven framework reformulates discrete hemispherical multipath error modeling into a continuous, coordinate-conditioned neural mapping function. Across extensive multi-week field evaluations conducted under contrasting physical conditions, the continuous modeling architecture substantially mitigated single-differenced carrier-phase residual errors across heterogeneous multi-constellation and multi-frequency signals. In three-dimensional position tracking, the framework reduced positioning root-mean-square errors from 7.8 millimeters to 4.7 millimeters at a slow-deformation landslide monitoring station and from 7.2 millimeters to 3.8 millimeters at a severely obstructed highway slope monitoring site. At the highly obstructed slope installation, the framework simultaneously increased the integer carrier-phase ambiguity fixing rate from 89.7 percent to 99.7 percent, demonstrating that continuous functional modeling can resolve critical signal distortions that conventional discrete lookup tables fail to mitigate.
In modern geodesy, autonomous vehicle guidance, structural deformation tracking, and hazard warning systems, satellite navigation infrastructure must deliver uninterrupted, millimeter-level coordinates under variable atmospheric and surrounding geometric conditions. While atmospheric delays such as ionospheric and tropospheric refractions, as well as satellite orbital anomalies and satellite clock drifts, can be largely mitigated or canceled out through differential processing techniques over short baselines, multipath effects persist stubbornly. Multipath phenomena arise when direct satellite radio waves interfere with indirect reflections rebounding from surrounding terrain, structures, or localized topography before reaching the antenna phase center. Because these secondary path reflections depend on the unique local physical environment, receiver antenna gain patterns, and constantly shifting satellite-reflector geometries, multipath errors exhibit highly nonlinear and non-stationary physical behavior. Standard double-differencing operations and stochastic parameter estimation models cannot eliminate these localized phase distortions, leaving residual phase errors that directly corrupt short-baseline real-time kinematic calculations and degrade integer ambiguity fixing reliability.
Engineering efforts to overcome multipath contamination have historically fallen into three primary methodologies: observation site optimization, hardware-level receiver filtering, and software-based algorithmic compensation. Site selection and antenna choke-ring designs can suppress localized ground reflections, but geodetic monitoring stations tracking active geological hazards, slopes, or infrastructure cannot choose benign surrounding topography and must remain rigidly fixed to vulnerable physical targets. High-end multi-element antenna arrays, specialized internal correlation processing loops, and integrated inertial measurement units escalate receiver hardware expenses, restricting widespread field deployment across low-cost geodetic sensor arrays. Consequently, software-driven algorithmic mitigation represents the most adaptable strategy for permanent continuous tracking stations. Algorithmic techniques include stochastic weighting adjustments and deterministic spatiotemporal mapping. In deterministic modeling, temporal sidereal filtering exploits orbit repetition periods to apply previous day phase residuals as subsequent corrections. Conversely, spatial multipath hemispherical maps discretize the local hemisphere into elevation and azimuth bins, populating fixed skyplot cells with historical residuals to form empirical lookup tables for incoming satellite trajectories.
Despite widespread adoption, conventional spatial multipath hemispherical maps suffer from fundamental representational bottlenecks. Traditional hemispherical maps rely on rigid spatial discretization grids, commonly configured at one-degree by one-degree resolutions, which inherently truncate spatial resolution and introduce step-like boundary discontinuities when satellite sightlines transition across adjacent bins. These discrete lookup bins cannot model continuous spatial gradations and lack a unified mathematical structure capable of jointly encoding multi-system and multi-frequency observations within a single shared model. Furthermore, while standard machine learning regressors such as multilayer perceptrons and gradient-boosted decision trees can fit complex nonlinear mappings, classical deep neural networks exhibit pronounced spectral bias. Under spectral bias, standard networks preferentially optimize for low-frequency global trends while severely suppressing high-frequency spatial variations. This frequency attenuation prevents conventional deep architectures from capturing the rapid, sharp localized phase fluctuations caused by localized obstruction edges, sharp reflection horizons, and steep terrain boundaries.
The newly developed INR-MHM architecture resolves these spatial and spectral limitations by adapting implicit neural representations from computer vision and three-dimensional coordinate modeling to geodetic signal processing. Instead of storing discrete residual averages across pre-allocated spatial bins or treating coordinates merely as auxiliary tabular predictors in sample-wise regressors, the framework parameterizes the entire local multipath field as a smooth, continuous coordinate function. The satellite line-of-sight unit direction vector serves as the primary spatial query coordinate, defined within a topocentric hemispherical coordinate system using elevation and azimuth geometry. By embedding the multipath error distribution within the continuous weights of a neural network, the system permits arbitrary coordinate querying across the entire upper hemisphere without being constrained by grid dimensions, interpolation boundaries, or artificial bin limits. Satellite constellation categories, operational transmission frequencies, and signal-quality telemetry provide supplementary contextual coordinates, enabling a single unified model to simultaneously represent disparate signals traversing the same physical environment.
To overcome neural spectral bias and resolve sharp spatial boundaries, the INR-MHM architecture incorporates multi-scale Fourier spatial encoding before feeding coordinates into deep network layers. The directional unit vector is mapped into a high-dimensional feature representation using three parallel Gaussian projection matrices governed by dyadic scale factors of 1.0, 2.0, and 4.0. The baseline scale of 1.0 captures broad, smooth geometric multipath variations across the open sky, while the higher scales of 2.0 and 4.0 sample the frequency domain on an expanded logarithmic variance, projecting low-dimensional coordinates through explicit sinusoidal basis functions. Crucially, these Gaussian projection matrices remain entirely fixed and non-trainable after initial generation from standard normal distributions, preventing gradient disruption while expanding the bandwidth of the effective neural tangent kernel. This mathematical projection transforms smooth geometric coordinates into high-dimensional sinusoidal features, enabling subsequent fully connected layers to approximate rapid spatial oscillations and sharp reflection boundaries without succumbing to low-frequency over-smoothing.
Simultaneously, heterogeneous observation characteristics are harmonized through dedicated embedding mechanisms and adaptive gating modules. Because discrete constellation identifiers and signal carrier frequencies do not possess a natural metric distance, each categorical attribute is transformed into a dense, continuous sixteen-dimensional latent embedding vector. Continuous signal-quality parameters, comprising normalized signal-to-noise ratio measurements and rolling standard deviations calculated over preceding twelve-epoch sliding windows, are projected through linear layers with Sigmoid Linear Unit activations into a thirty-two-dimensional vector. These embedded components are concatenated with the Fourier-encoded spatial vector and projected into a unified latent feature space of 256 dimensions. To dynamically adjust the influence of these diverse inputs under changing tracking conditions, an adaptive context-gated fusion module employs a two-layer bottleneck network with dimensionality reduction, nonlinear activation, dropout, and dimension restoration. This bottleneck produces a 256-dimensional gating vector that recalibrates latent channels via element-wise multiplication prior to feeding a lightweight nonlinear regression head that outputs single-differenced phase corrections.
Extracting the underlying multipath signals from raw field data requires rigorous preprocessing to isolate physical phase distortions from receiver noise and baseline parameters. In short-baseline real-time kinematic processing, double-differenced carrier-phase observation equations eliminate satellite and receiver clock errors while mitigating residual atmospheric propagation delays, leaving residual vectors dominated by multipath fluctuations and stochastic measurement noise. Because double-differenced residuals represent satellite pairs rather than individual satellite links, researchers converted double-differenced residuals into satellite-specific single-differenced residuals by imposing an elevation-weighted or unit-weighted zero-mean mathematical constraint to resolve the intrinsic rank deficiency of the differencing matrix. Once single-differenced carrier-phase residuals were recovered for each tracked satellite, high-frequency white measurement noise was eliminated using a Savitzky-Golay filter. This filtering process smoothed raw high-frequency receiver noise while preserving underlying multipath amplitudes and localized phase curvatures without inducing artificial signal attenuation or temporal phase shifts, isolating millimeter-level phase targets for model supervision.
The empirical validation of INR-MHM was conducted using extensive observational datasets gathered across independent 21-day testing windows at two geodetic stations exhibiting contrasting physical topographies in Gansu Province, China. Station HF23, situated on the active Heifangtai landslide, represented a relatively open observation environment undergoing slow, millimeter-level horizontal tectonic displacement, monitored using a low-cost u-blox F9P receiver tracking GPS, BeiDou, and Galileo signals. In contrast, station GZY12, positioned along a steep highway slope in Shichuan, Lanzhou, represented a severely obstructed environment marked by prominent terrain blockages, degraded southern satellite skyplots, and severe reflection hazards, monitored via a BDStar U6 receiver tracking GPS, BeiDou, Galileo, and GLONASS signals. Training was performed across seven consecutive days to provide wide spatial coverage across the local sky hemisphere, followed by two validation days for learning-rate scheduling and early stopping, and twenty-one independent testing days to evaluate temporal generalization against conventional grid-based hemispherical maps, second-order trend-surface hemispherical maps, and gradient-boosted decision trees.
Observation-domain evaluations confirmed that continuous functional modeling provided superior fidelity when reconstructing carrier-phase multipath residuals over multi-week testing periods. Across individual satellite arcs, power spectral density analyses demonstrated that INR-MHM reproduced residual power spectra accurately within the critical 0.001 to 0.1 Hertz frequency band, where conventional grid maps and trend-surface approximations exhibited pronounced spectral attenuation and smoothing. Over the entire 21-day testing period, INR-MHM achieved lower mean daily single-differenced carrier-phase residual root-mean-square errors than all baseline models across every evaluated constellation and signal frequency. Relative to uncorrected raw observations, the proposed model achieved residual root-mean-square reductions ranging between 34.7 percent and 72.2 percent at the open HF23 landslide monitoring station. At the severely obstructed GZY12 highway slope installation, single-differenced residual reductions spanned 52.4 percent to 61.0 percent, proving that the continuous coordinate mapping maintained stable correction performance without suffering from the edge degradation seen in discrete grid models.
In the positioning domain, applying continuous multipath corrections directly into short-baseline real-time kinematic processing pipelines produced substantial accuracy gains across all coordinate components. At station HF23, where uncorrected observations yielded fixed-solution root-mean-square errors of 2.3 millimeters in east, 3.6 millimeters in north, and 6.6 millimeters in up, INR-MHM reduced component errors to 1.5 millimeters, 1.8 millimeters, and 4.1 millimeters, respectively, driving the overall three-dimensional positioning root-mean-square error from 7.8 millimeters down to 4.7 millimeters. Performance gains were even more striking at the obstructed GZY12 slope monitoring station, where severe terrain multipath degraded uncorrected coordinates to 2.3 millimeters east, 3.3 millimeters north, and 6.0 millimeters up. Applying INR-MHM reduced these errors to 1.2 millimeters, 1.2 millimeters, and 3.3 millimeters, lowering the overall three-dimensional root-mean-square error from 7.2 millimeters to 3.8 millimeters while boosting the carrier-phase ambiguity fixing success rate from 89.7 percent to 99.7 percent.
Rigorous step-by-step ablation analyses confirmed the distinct operational roles played by the framework’s core architectural components. Comparing baseline coordinate regression models without Fourier mapping against Fourier-encoded architectures demonstrated that multi-scale Fourier spatial encoding served as the primary driver of residual root-mean-square error reduction, especially at the highly obstructed station where fine-scale multipath variations dominated the residual spectrum. While incorporating context-gated feature fusion yielded smaller incremental reductions in residual root-mean-square values than Fourier mapping, the gating mechanism provided essential sample-dependent modulation that adaptively harmonized heterogeneous spatial geometry, system-frequency identifiers, and dynamic signal-to-noise ratios. The study authors noted that while seven days of training provided robust spatial coverage for multi-week generalization, seasonal environmental changes, localized vegetation growth, snow accumulation, and structural alterations will eventually modify local multipath fields over extended durations, highlighting periodic model updating and continual learning strategies as valuable avenues for future geodetic research.
Subject of Research: Continuous hemispherical modeling and mitigation of GNSS carrier-phase multipath errors using implicit neural representations
Article Title: INR-MHM: an implicit neural representation framework for continuous hemispherical modeling of GNSS multipath errors
Article References: Li, D., Du, Y., Huang, G., & Zhang, Q. (2026). INR-MHM: an implicit neural representation framework for continuous hemispherical modeling of GNSS multipath errors. Satellite Navigation, 7(1), Article 45. https://doi.org/10.1186/s43020-026-00234-8
Image Credits: AI Generated
DOI: 10.1186/s43020-026-00234-8
Keywords: Global Navigation Satellite System, Multipath Error Mitigation, Implicit Neural Representation, Multipath Hemispherical Map, Carrier-Phase Residuals, Multi-Scale Fourier Encoding, Spectral Bias, Context-Gated Feature Fusion, Real-Time Kinematic Positioning, Ambiguity Resolution, Single-Differenced Residuals, Double-Differenced Processing, Savitzky-Golay Filter, Landslide Deformation Monitoring, Highway Slope Monitoring, Multi-Frequency GNSS, Multi-Constellation Navigation, Signal-to-Noise Ratio
Cite Scienmag News
Cassandra Pierce. (November 22, 2026). Implicit Neural Representation Models Continuous GNSS Multipath Errors to Boost Positioning Accuracy. Scienmag. https://scienmag.com/implicit-neural-representation-models-continuous-gnss-multipath-errors-to-boost-positioning-accuracy/
Cassandra Pierce. "Implicit Neural Representation Models Continuous GNSS Multipath Errors to Boost Positioning Accuracy." Scienmag, 22 November 2026, https://scienmag.com/implicit-neural-representation-models-continuous-gnss-multipath-errors-to-boost-positioning-accuracy/. Accessed 22 November 2026.
Cassandra Pierce. "Implicit Neural Representation Models Continuous GNSS Multipath Errors to Boost Positioning Accuracy." Scienmag. November 22, 2026. https://scienmag.com/implicit-neural-representation-models-continuous-gnss-multipath-errors-to-boost-positioning-accuracy/








