Jianping Chen and Yang Gao, Department of Geomatics Engineering, University of Calgary, Canada

View Abstract Sign in for premium content


In addition to ionosphere remote sensing applications using GNSS, the availability of high precision ionospheric corrections is essential for precise positioning using low-cost single-frequency GNSS receivers. The precise ionospheric corrections can also improve the convergence of precise positioning using dual-frequency GNSS receivers. The International GNSS Service (IGS) is providing Real-Time Global Ionosphere Maps (RT-GIM) which are the combined products from the IGS real-time ionosphere centers. With the latest interpolation improvements, the RT-GIM accuracy is close to that of the IGS rapid GIMs. However, the prediction methods still need globally distributed real-time GNSS stations. This paper investigates the development of a regional precise ionospheric model based on a sequence-to-sequence long short-term memory (LSTM) deep learning method to predict ionospheric vertical delay maps for real-time applications. The Recurrent Neural Network (RNN) model was derived from feedforward neural networks to deal with complex temporal problems. The LSTM model is a special case of the powerful RNN model with the ability to overcome the gradient varnishing issue from the traditional RNN. In this research, the IGS rapid GIMs are used as prediction bases so real-time GNSS datalinks are not needed. Various time windows are selected based on different Kp-index to represent different solar activity strengths over a region in North America. The ionospheric corrections estimated from GNSS stations within the testing region are used to verify the model performance. The training data for the testing region is retrieved from the IGS Global Ionosphere Map (GIM) product, including 90 days before the prediction day. Comparing the ionospheric predictions over 24 hours using the proposed model with the IGS Final Global Ionosphere Map (GIM) product indicates that the model has a prediction accuracy at a root-mean-square error (RMSE) of 0.8 TECU and about 2 TECU for one week. Since the training is based on the IGS Rapid GIM product and no more real-time GNSS data is needed, the model can provide improved cost-effective regional ionospheric services with high stability.