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

Smartphone global navigation satellite system (GNSS) positioning is degraded by low-cost antennas, limited receiver hardware, multipath propagation, and noisy code pseudorange observations. Existing correction methods often improve stochastic weighting, estimate coordinate-domain corrections, or smooth receiver trajectories, but they rarely estimate how each satellite contributes to the horizontal position error while preserving line-of-sight (LOS) geometry. This study presents a random-forest-assisted geometry-aware correction method that combines satellite-wise LOS projection error estimation with exponential temporal weighted least squares (Temporal WLS). The horizontal error between the smartphone National Marine Electronics Association (NMEA) solution and the F9P reference position is projected onto each satellite LOS direction and used as the learning target. A random forest model is trained using 26 smartphone GNSS features, including geometry, signal strength, code-derived variation, uncertainty, automatic gain control, and state flags. The predicted LOS errors are fused with satellite geometry through epoch-wise WLS and Temporal WLS. In same-session front-70/back-30 validation, the horizontal root mean square (RMS) error decreased from 2.747 m to 1.033 m. Excluding one suspected non-co-located reference session further reduced the RMS error from 2.867 m to 0.362 m.