Random-Forest-Based Smartphone GNSS Position Correction Using Satellite-Wise LOS Projection Error Estimation and...

Area of Science:

• Geomatics Engineering
• Satellite Navigation Systems
• Machine Learning Applications

Background:

• Smartphone Global Navigation Satellite System (GNSS) positioning suffers from signal degradation due to low-cost hardware and environmental factors like multipath propagation.
• Current correction methods often lack the ability to precisely quantify individual satellite contributions to horizontal position errors while maintaining line-of-sight (LOS) geometry.

Purpose of the Study:

• To develop a geometry-aware correction method for smartphone GNSS positioning.
• To estimate satellite-specific line-of-sight (LOS) projection errors.
• To improve the accuracy of horizontal position estimation in challenging GNSS environments.

Main Methods:

• A random-forest model was trained using 26 diverse smartphone GNSS features.
• Satellite-wise LOS projection errors were estimated using the horizontal error between smartphone (NMEA) and reference (F9P) positions.
• Exponential temporal weighted least squares (Temporal WLS) was employed to fuse predicted LOS errors with satellite geometry.

Main Results:

• The horizontal root mean square (RMS) error was reduced from 2.747 m to 1.033 m using same-session validation.
• Excluding a potentially non-co-located reference session further improved accuracy, reducing RMS error from 2.867 m to 0.362 m.

Conclusions:

• The proposed random-forest-assisted, geometry-aware method effectively improves smartphone GNSS positioning accuracy.
• This approach offers a novel way to correct for satellite-specific errors, enhancing navigation performance in real-world conditions.