TransXLT: A novel ZTD prediction method with SASR-based data reconstruction

Area of Science:

• Geodesy and Satellite Navigation
• Atmospheric Science and Meteorology
• Artificial Intelligence and Machine Learning

Background:

• Traditional Zenith Tropospheric Delay (ZTD) models struggle with accuracy during complex weather and data gaps.
• Global Navigation Satellite System (GNSS) data are crucial for precise positioning but are affected by tropospheric delays.

Purpose of the Study:

• To develop an advanced model for accurate ZTD estimation, addressing limitations of existing methods.
• To enhance ZTD prediction accuracy by integrating diverse data sources and robust data imputation techniques.

Main Methods:

• A novel transformer-xLSTM (TransXLT) model was developed, combining spatial-temporal information from GNSS, ERA5, and GPT3.
• A sparse attention-based time series reconstruction (SASR) method was employed to handle missing GNSS data.

Main Results:

• SASR reduced Mean Absolute Error (MAE) by 24.5% and training Root Mean Square Error (RMSE) by 15.1% under significant data loss.
• The TransXLT model achieved an average RMSE of 8.13 mm across six sites, outperforming benchmarks by up to 76.54%.
• The model demonstrated robustness across different latitudes, altitudes, and seasons.

Conclusions:

• The TransXLT model significantly improves ZTD estimation accuracy for GNSS applications, especially under challenging conditions.
• The integration of advanced AI techniques and multi-source data offers a promising direction for precise atmospheric parameter estimation.