How LightGBM Boosts Urban GNSS Accuracy by Detecting NLOS Errors

Introduction

With the rapid growth of intelligent transportation and autonomous driving, the positioning accuracy and reliability of Global Navigation Satellite Systems (GNSS) become critical. In dense urban areas, buildings and vehicles cause signal reflections and refractions, leading to non‑line‑of‑sight (NLOS) errors that can reach hundreds of meters, severely limiting GNSS applications.

Existing approaches—signal processing optimization, multi‑sensor fusion, 3D‑map assistance, and traditional machine‑learning methods—suffer from subjective sample labeling, insufficient feature analysis, and an imbalance between real‑time performance and accuracy.

Key Innovations

The collaboration between Baidu Maps and Wuhan University produced a paper titled A reliable NLOS error identification method based on LightGBM driven by multiple features of GNSS signals , achieving breakthroughs in dataset construction, feature selection, and model design.

Objective labeling using a fused fisheye‑camera + inertial‑navigation system, eliminating the subjectivity of residual‑based labeling.

Construction of a multi‑dimensional GNSS feature set, quantifying feature‑label correlations and overcoming the limitations of single‑feature methods.

Adoption of an optimized LightGBM classifier that balances high precision with real‑time inference suitable for vehicular scenarios.

Technical Details

1. NLOS Sample Annotation Technique

Data were collected with an Entaniya Fisheye M12 280° camera and a SPAN ISA‑100C inertial navigation unit. The fisheye images were undistorted using the Kannala‑Bandt model, then processed through edge detection, contour completion, and watershed segmentation to extract the sky region. Satellite positions were projected onto the camera image plane using GNSS ephemerides and vehicle pose, enabling direct visual labeling of LOS/NLOS signals.

Sky‑region extraction flow: preprocessing → Canny edge detection → contour completion → watershed segmentation, resulting in accurate sky masks for labeling.

2. Feature Engineering

Six GNSS‑derived features sensitive to NLOS were extracted:

SNR (Signal‑to‑Noise Ratio) : reflects signal strength; LOS signals concentrate between 40‑50 dB.

EA (Elevation Angle) : correlates with blockage probability; low angles (≤20°) are predominantly NLOS.

PRC (Pseudorange Consistency) : validates pseudorange reliability using Doppler observations.

PC (Phase Consistency) : captures phase jumps.

CMC (Code‑Carrier Difference) : amplifies NLOS error after double‑difference and epoch‑difference corrections.

MP (Multipath Observation) : quantifies pseudorange bias via dual‑frequency combination.

All features were Z‑score normalized to remove unit disparities.

3. Model Selection and Core Optimizations

LightGBM was chosen over XGBoost and CatBoost for its histogram‑based split algorithm, leaf‑wise growth with depth constraints, gradient‑based one‑side sampling (GOSS), and exclusive feature bundling (EFB). These optimizations reduced training complexity, prevented over‑fitting, and achieved inference times ≤100 µs, meeting vehicular real‑time requirements.

4. Training and Inference Pipeline

Training employed 5‑fold cross‑validation, Z‑score standardized features, and binary cross‑entropy loss. The best model (F1‑score 0.9232) was selected for deployment. During online inference, the six GNSS features are extracted, standardized, fed to the model, and the output determines LOS (probability ≥ 0.5) or NLOS (probability < 0.5), after which NLOS observations are discarded from positioning calculations.

Experimental Results

Experiment 1 – Impact of Feature Quantity

Increasing the number of features from 1 to 6 raised the F1‑score by 36.4 %, precision by 15.3 %, and reduced misclassifications by 67.2 %. Elevation Angle proved the most discriminative feature.

Experiment 2 – Model Accuracy and Efficiency Comparison

LightGBM achieved the highest precision (92.55 %) and the lowest NLOS miss‑detection rate (5.49 %) compared with XGBoost and CatBoost, while training time (25.49 s) and inference time (0.97 s) were substantially lower than XGBoost and comparable to CatBoost.

Experiment 3 – Positioning Performance After NLOS Removal

Removing NLOS observations reduced RMS errors in dual‑frequency SPP by 47.8 % (East), 56.7 % (North), and 36.7 % (Up). In single‑frequency scenarios, Up‑direction error decreased by 45.2 %. Satellite visibility remained sufficient (12‑18 LOS satellites), and PDOP improved from 3‑5 to 2‑3.

Conclusion

The proposed GNSS‑based NLOS identification method, driven by multi‑feature LightGBM, delivers objective dataset construction, comprehensive feature analysis, and an efficient high‑accuracy model, leading to significant positioning improvements in complex urban canyons.