LF-GNSS: Towards More Robust Satellite Positioning with a Hard Example Mining Enhanced Learning-Filtering Deep Fusion Framework

Global Navigation Satellite System (GNSS) is essential for unmanned systems and location-based technologies, providing critical geospatial information for navigation and situational awareness. However, GNSS performance degrades significantly in urban environments due to Non-Line-of-Sight (NLOS) conditions and multipath effects. To address these challenges, we propose LF-GNSS, a learning-filtering deep fusion framework with satellite-aware feature modeling, which integrates deep learning with classical filtering techniques while preserving model interpretability. The framework incorporates a DOP-based representation that encodes each satellite’s geometric influence, enabling adaptive measurement weighting within the filter. In addition, a dynamic hard example mining strategy is employed to emphasize challenging samples during training, further enhancing robustness under degraded conditions. LF-GNSS has been extensively validated on both public and private datasets, showing consistent improvements in positioning accuracy over traditional methods and existing learning-based approaches. Runtime evaluations on embedded platforms further indicate the framework’s feasibility for real-time deployment in practical navigation systems. To foster further research and development, we will publicly release our code at https://github.com/GarlanLou/LF-GNSS and make available a dataset at https://github.com/GarlanLou/LF-GNSS-Dataset.