Detection and Mitigation of NLOS and Multipath Effects Using Multi-Task Learning

Ellarizza Fredeluces and Nobuaki Kubo

Abstract:	In this study, we use the correlation values as input to ResNet18 and do multi-task learning to classify NLOS and LOS, and estimate multipath error. Collected data from three sites around Tokyo station in Tokyo, Japan were post-processed in a software receiver to get the correlation values. DGPS method is used to isolate multipath error of target satellite and use that as classification and regression labels. After training, we were able to get 98.05% accuracy, while precision, recall, and fscore values for two categories differ by 3-4%. For multipath error estimation, total mean absolute error and RMSE are 30.8m and 75.96m, respectively. LOS RMSE is 7.25m while NLOS RMSE is 193.23m. The model performed better for LOS category than NLOS category due to data imbalance. In the positioning domain, the proposed method’s 2DRMS has decreased by 7.2% compared with normal DGPS. The quality of raw pseudoranges were improved when estimated multipath error is added. This is shown by higher percentage of output in RTKLIB.
Published in:	Proceedings of the 2025 International Technical Meeting of The Institute of Navigation January 27 - 30, 2025 Hyatt Regency Long Beach Long Beach, California
Pages:	906 - 913
Cite this article:	Fredeluces, Ellarizza, Kubo, Nobuaki, "Detection and Mitigation of NLOS and Multipath Effects Using Multi-Task Learning," Proceedings of the 2025 International Technical Meeting of The Institute of Navigation, Long Beach, California, January 2025, pp. 906-913. https://doi.org/10.33012/2025.19967
Full Paper:	ION Members/Non-Members: 1 Download Credit Sign In