An Enhanced Method for NLOS Signal Detection in Urban Environments based on Stacking Ensemble Learning for Smartphone Positioning

Global Satellite Navigation System (GNSS) has been widely used in our daily life, due to its high-precision positioning ability. However, in urban environments, GNSS signals are prone to obstructions from tall buildings, leading to Non-Line-of-Sight (NLOS) errors and a significant decline in positioning accuracy. Machine learning (ML) techniques for NLOS detection has emerged as a significant research hotspot, thanks to its advantages, such as requiring no hardware modifications, achieving high accuracy, and offering practical applicability, etc. However, the existing research on ML-based NLOS detection usually relies on single models, which are prone to becoming trapped in local optima during the optimization process. To overcome this limitation, we propose a GNSS NLOS detection method based on the Stacking Ensemble Learning (SEL) model. Initially, the fisheye camera is utilized to generate NLOS labels, and five GNSS signal features are extracted: elevation angle, carrier-to-noise ratio (C/N₀), measurement residual, C/N₀ rate of change, and pseudorange standard deviation. Subsequently, the SEL model is designed with a two-layer structure. The first layer consists of basic ML classification models, including Convolutional Neural Network (CNN), Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and Support Vector Machine (SVM). The second layer employs Logistic Regression (LR) as the meta-learning model to integrate the outputs from the first layer. Finally, the trained SEL model processes the five GNSS signal features in real time for detecting smartphone GNSS NLOS signals, and incorporates a weighted model for SPP positioning. Several smartphone-based vehicle experiments were conducted in different urban areas of Wuhan, China, to validate the effectiveness of the proposed method. Experimental results demonstrate that the SEL method achieves GNSS NLOS detection accuracies exceeding 90%, with detection performance improvements ranging from 15.6% to 32.7%, compared with the single ML methods such as CNN, GBDT, RF, and SVM. Furthermore, the SEL method enhances 3D positioning accuracy, with improvements ranging from 26.7% to 39.6%. Particularly, in dense urban canyon areas, the vertical positioning accuracy is improved by up to 73.1%, effectively mitigating the impact of NLOS signals. This method requires no additional improvements to low-cost receiver hardware, thus offers potential for widespread application across various GNSS terminals, and provides new ideas for navigation and positioning in smart cities.