Session B3: AI and Machine Learning for GNSS Signal Processing, Error Modeling, and Threat Detection

Deep Learning Frameworks for Ionosphere Modeling: Comparison under Multiple Datasets
Yifei Chen, Chengcheng Song, School of Instrumentation and Opto-Electronic Engineering, Beihang University, Beijing, P.R.China; Zejian Ding, Department of Computer Science, City University of Hong Kong, Hong Kong, P.R. China; Yang Liu, School of Instrumentation and Opto-Electronic Engineering, Beihang University, Beijing, P.R.China; Tianmushan Laboratory, Beihang University, Hangzhou, P.R.China; Marconi Lab, Science Technology and Innovation Section, Abdus Salam International Centre for Theoretical Physics
Location: Windsong 7-8

The ionosphere is a part of Earth's atmosphere characterized with layers of ionized gas, containing large amounts of free electrons and positive ions. Due to its properties, the ionosphere plays a crucial role as a medium for radio wave propagation, affecting the performance of technologies like the Global Navigation Satellite System (GNSS). One key parameter for describing the ionosphere's electron density is the Total Electron Content (TEC). TEC represents the integral of electron density along the vertical direction over a unit area, commonly quantified as the total number of electrons per unit area. Measuring TEC provides valuable information on the structure and variability in the ionosphere and is very significant for those research areas such as space communication, satellite navigation, and space weather monitoring. Accurate prediction of TEC aids in optimizing the performance of GNSS systems, improving the accuracy and stability of positioning, identifying and mitigating electromagnetic interference, and supporting the monitoring and forecasting of space weather phenomena.
The recent advances in artificial intelligence, especially deep learning techniques, shed lights on modeling and prediction for ionosphere TEC with multiple observations. A lot of works focuses on deep learning ionosphere modeling based on the strong nonlinear fitting capabilities and superior feature extraction performance, with core concern of precise forecasting capability for the models. For the TEC observations, global ionospheric map (GIM) products provided by the International GNSS Service (IGS) become the most used dataset for training and testing, and TEC modeling accuracy closely relies on those GIM products issued by different IGS data analysis centers.
To address the problem, this work takes concerns of different GIM products from four data sources, namely the IGS final data, UPC data, JPL data and the data generated by Beihang University. A unified spatial temporal TEC prediction task was constructed, based on those four GIM datasets. The global TEC morphology was forecasted one day ahead with seven days of historical data. In the task, several deep learning ionosphere TEC modeling frameworks were established, including two-layer/three-layer ED-ConvGRU, iTransformer, MLPMultivariate, and SOFTS. The testing experiment was concentrated on two strong geomagnetic storms in 2015, the March 17 storm and June 23 storm. Experimental results reveal significant performance variations across models and datasets. For IGSG data, a comparison study evaluated performances of ED-ConvGRU against iTransformer, MLPMultivariate, and SOFTS during two selected geomagnetic storms (Event A: March 17; Event B: June 23). It shows that 3-layer ED-ConvGRU achieves MAE=3.72/RMSE=5.83 in Event A (21.0%/9.8% improvement over MLPMultivariate) and MAE=2.69/RMSE=4.08 in Event B (27.6%/19.0% improvement). Notably, ED-ConvGRU exhibits the lowest RMSE fluctuation (18.7% lower than iTransformer) in Event B. The dual-event analysis confirms ED-ConvGRU's superior performances in accuracy and stability under strong geomagnetic storms, when global TEC facing great perturbations. Similar experiments were conducted with other three data sources.
The significance of this work lies in addressing a critical gap in global TEC temporal prediction. By utilizing multiple deep learning methods, it offers new perspectives and solutions for researchers and practitioners in the field. First, this study systematically evaluates the effectiveness of several deep learning models in global TEC prediction under strong geomagnetic storms. We compared the models’ performances across various strong geomagnetic storms. Second, we fully consider the variability and influences of different data sources. In summary, this study not only improves the accuracy of TEC prediction under geomagnetic storms but also introduces new methodologies and perspectives to the field. It holds good implications for enhancing positioning accuracy, mitigating electromagnetic interference, and advancing space weather monitoring and forecasting.