ML-based Multipath Modeling for Precise Point Positioning

Abstract

Permanently installed GNSS reference stations form the backbone of modern geodetic infrastructure, supporting applications that range from structural deformation monitoring and national coordinate reference frames to telecommunications timing and precision agriculture. At these sites, Precise Point Positioning (PPP) can achieve centimetre-level accuracy of the reference station position using only a single receiver and external precise orbit and clock products, without the need for a nearby base station. After orbital, clock, and atmospheric errors have been removed; however, multipath interference — caused by satellite signals reflecting off surrounding structures — remains the dominant residual error source and directly limits both positioning accuracy and convergence speed.

This thesis evaluates six multipath mitigation methods using two independent datasets. Theprimary dataset was collected over 31 days at a u-blox AG rooftop antenna in Thalwil, Switzerland; models are trained on Days 1–10 and evaluated on Days 11–31 (21-day test window). Cross-site validation is performed using three selected test days at the International GNSS Service reference station at Curtin University in Perth, Australia, to validate the trained models via cross-site testing without any pipeline modification. Two classical mitigation strategies — the Multipath Hemispherical Map (MHM) and the Sidereal Filter (SF) — are compared against four machine learning models: XGBoost, LightGBM, CatBoost, and Multilayer Perceptron (MLP). All methods are applied to ionosphere-free post-fit pseudorange and carrier-phase residuals from GPS and Galileo PPP processing and evaluated using the root-mean-square (RMS) reduction.

On the Thalwil dataset, tree-based machine learning models (XGBoost, LightGBM, CatBoost) achieve average pseudorange RMS reductions of 36.9–37.3% for the combined GPS/Galileo constellation, compared to 27.6% for MHM and 29.5% for SF. On Galileo pseudorange residuals specifically, XGBoost reaches 41.9%, compared to 27.3% for the Sidereal Filter — the weakest method on this configuration. The disparity arises because the SF reference day for Galileo is ten days prior, aligned with the 10-day orbital repeat period, and the environmental gap between reference and test day introduces atmospheric and thermal variability that degrades correction quality. On GPS carrier-phase residuals, by contrast, the SF (30.4%) performs comparably to the tree models (XGBoost 31.1%, LightGBM 32.0%) — the one configuration where a one-day temporal lag transfers cleanly. MLP trails the tree cluster by 3–4 percentage points on pseudorange and 4–5 percentage points on carrier-phase. At the positioning level, XGBoost correction reduces GPS-only convergence time by 50.8% (from 32.5 to 16.0 minutes) and improves the post-convergence three-dimensional position RMS by 29.2% for the GPS+Galileo configuration (from 17.4 to 12.3 mm).

On the Curtin University dataset, the pseudorange residual standard deviation after XGBoost correction (1.06–1.16 m across three test days) is comparable to that observed at Thalwil (1.05–1.15 m), despite the two sites having different multipath environments, confirming that the trained pipeline transfers to an independent site without modification.