Intelligent GNSS Satellite Visibility Classification in Urban Areas: A Deep Learning Approach with Interpretation

Accurate and reliable GNSS solutions are essential for the development of intelligent transportation systems. However, GNSS signals can be easily blocked by buildings in urban areas, resulting in the sole reception of the reflected signal with large errors, namely non-line-of-sight (NLOS) receptions. Thus, it is necessary to classify the visible satellite measurements from NLOS receptions before conducting positioning. This paper aims to design a Transformer-based deep learning network to utilize the spatial correlations between satellites for their visibility classification. The proposed method achieves about 89 percent classification accuracy in validation and test data. By exploring the spatial correlation between satellites in the attention matrix of Transformer, we reveal the mechanism of deep learning network on satellite visibility classification.