NLOS signal identification in urban canyons based on a hybrid LSTM_LightGBM model

In urban environments, global navigation satellite system (GNSS) signals are prone to obstruction and reflection caused by surrounding structures, leading to the reception of non-line-of-sight (NLOS) signals. This effect significantly impairs the accuracy and reliability of satellite-based positioning. This paper proposes an NLOS signal identification method based on a hybrid model that integrates long short-term memory (LSTM) networks with light gradient boosting machine (LightGBM). Firstly, the relationships between signal types and seven key features – including pseudorange standard deviation, carrier phase standard deviation, Doppler frequency standard deviation, carrier phase lock-time counter, and carrier-to-noise ratio – are systematically analysed. Then, the LSTM network is employed to extract deep temporal features from the GNSS signal sequences. These learned representations are subsequently fused with the original observation data to construct a comprehensive hybrid feature space. Finally, a LightGBM classifier is utilised to perform signal classification. Experimental results demonstrate that the proposed method achieves an accuracy of 95.03 % and an F1-score of 93.96 % on the test dataset, showing significant improvements over baseline models such as random forest, K -nearest neighbours, gradient boosted decision trees, standalone LSTM and stand-alone LightGBM. Notably, the model maintains an accuracy of 95.90 % even under class-imbalanced conditions, confirming its robustness and generalisation capability in complex urban canyon scenarios.