Reconstructing missing global positioning data with zero-shot large language models

This study focuses on the reconstruction of missing GPS trajectory data. The principal issue relates to restoring geospatial coordinates in the absence of large volumes of labeled data and under conditions where conventional spatial-temporal models demonstrate limited generalization capabilities. This paper proposes a large language model-based approach to address the reconstruction task without requiring prior training on specialized datasets. To reduce dependence on domain-specific features, the focus was on optimizing data preprocessing and constructing effective prompts. Three coordinate representations have been explored: original degree-based values (using the VPAIR dataset), the Earth-Centered, Earth-Fixed (ECEF) system, and ECEF coordinates shifted relative to the starting point of the trajectory. Experimental results show that using centered ECEF coordinates reduces the mean absolute error (MAE) by 51–59% for both latitude and longitude compared to other representations. Conversion to the ECEF system also demonstrates selective advantages in latitude reconstruction. To mitigate the instability of autoregressive prediction, a multi-iteration reconstruction strategy with result aggregation has been implemented. The open-source model LLaMA 3.2 achieved the highest accuracy (MAE: 36.57 for latitude and 52.14 for longitude), outperforming both other open models and the commercial GPT-4o. The proposed approach can be considered a viable post-processing tool, particularly in missions involving unmanned aerial vehicles or other mobile platforms where part of the GPS data has been lost during acquisition