Enhancing GNSS Localization in Urban Canyons with a Hybrid CNN-Autoencoder Approach to LOS/NLOS Classification

Non-Line-of-Sight (NLOS) reception is acknowledged as a primary source of positioning error in Global Navigation Satellite System (GNSS) applications, particularly within dense urban settings. When the direct satellite path is obstructed by buildings and other structures, receivers must depend on the reflected signals. This reliance results in significant pseudorange biases and a diminished localization accuracy. Methods based on learning have been developed to classify signals as Line-of-Sight (LOS) or NLOS using signal features; however, many approaches depend on external sensors, heuristic thresholds, or hand-crafted features, restricting their scalability and adaptability. A hybrid deep learning (DL) framework is proposed to address these constraints, integrating a Convolutional Neural Network (CNN) with an Autoencoder to improve LOS/NLOS signal classification utilizing raw GNSS measurements. The model is trained and evaluated using the UrbanNav dataset, which offers dual-frequency multi-GNSS observations from urban canyons in Hong Kong and Tokyo. The training process encompasses feature normalization, class balance, and structured learning, with optimization conducted through the Adam optimizer and dynamic learning rate scheduling. It has been demonstrated through experimental results that the proposed approach achieves a global classification accuracy of 99.90%. For LOS signals, an accuracy of 99.95% is recorded, along with a precision, recall, and F1 score of 100%. In contrast, NLOS signals are accurately captured with 100% in all evaluation criteria. These outcomes highlight the robustness and precision of the proposed method, rendering it highly suitable for applications necessitating reliable urban positioning, such as autonomous driving, UAV localization, and pedestrian navigation.