Adaptive Extended Kalman Filtering with LSTM-Based Signal Quality Assessment for GNSS Precise Point Positioning

Multipath and non-line-of-sight (NLOS) signal reception remain major error sources in Global Navigation Satellite System (GNSS) Precise Point Positioning (PPP), particularly in urban environments where satellite signals are frequently reflected, diffracted, or obstructed. This paper presents a Long Short-Term Memory (LSTM)-based multipath detection and adaptive Extended Kalman Filter (EKF) weighting framework to improve PPP positioning accuracy. A scenario-based multi-constellation GNSS observation dataset is developed to model line-of-sight (LOS), multipath, and NLOS signal conditions under varying environmental scenarios. From each GNSS observation, six multipath-sensitive features are extracted: signal-to-noise ratio (SNR), elevation angle, code residual, ionospheric delay, delta code-minus-carrier (∆CMC), and code rate consistency (CRC). The extracted features are arranged into temporal sequences and processed by an LSTM network to estimate LOS, multipath, and NLOS probabilities. These probabilities are converted into a signal quality score and used to adaptively modify the observation covariance in a simplified PPP/EKF positioning framework. The proposed method is evaluated against a standard EKF using fixed measurement covariance. Simulation results show that the LSTM-adaptive EKF approach reduces horizontal RMSE from 4.211 m to 2.106 m and 3D RMSE from 8.362 m to 4.090 m, corresponding to improvements of 50.00\% and 51.09\%, respectively. The results indicate that temporal deep learning can provide useful observation-quality information for multipath-aware adaptive weighting in GNSS PPP.