Spatiotemporal Deep Learning Network for High­-latitude Ionospheric Scintillation Forecasting

Abstract:

Global navigation satellite system (GNSS) signal is known to experience scintillation effect during space weather activities. Ionospheric scintillation refers to rapid amplitude and/or phase fluctuations caused by signal propagation through ionospheric plasma irregularities. The occurrence of scintillation may impair the receiver tracking loop, resulting in enhanced noise and discontinuity in measurements and disruption of services. Therefore, it is important to forecast the occurrence of scintillation such that appropriate measures can be taken to alleviate the potential impact. The complex physics behind the generation and evolution of ionospheric irregularities makes it very challenging to forecast scintillation occurrence based on physical models. Thus, data-driven methods and machine learning techniques have been used as alternative approaches in recent years.

In this study, we apply a spatiotemporal deep learning (STDL) network [1] to forecast the GNSS signal phase scintillation at high latitudes. Similar to existing forecasting methods [2], [3], the network utilizes local features such as historical measurements from the GNSS target station and external features that are potential drivers of the scintillation occurrence including solar wind parameters, geomagnetic activity indices, etc. In addition, we also use global features such as historical measurements from surrounding GNSS receivers to provide spatial information. Rather than naively dumping all the features into a conventional machine learning algorithm (e.g. Logistic Regression), we use the STDL network to adaptively incorporate spatial and temporal information.

The results demonstrate that the STDL network adaptively utilizing local, global, and external features can achieve a 0.94 on the score of receiver operating characteristics (ROC) area under curve (AUC). The STDL performance is better than several other machine learning algorithms including logistic regression, fully connected neural network, and long short-term memory (LSTM). This presentation will discuss the STDL architecture, feature selections, training and testing data sets, and detailed performance assessment.

[1] Y. Liang, S. Ke, J. Zhang, X. Yi, and Y. Zheng, “GeoMAN: Multi-level Attention Networks for Geo-sensory Time Series Prediction,” in Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, Stockholm, Sweden, Jul. 2018, pp. 3428–3434, doi: 10.24963/ijcai.2018/476.

[2] R. M. McGranaghan, A. J. Mannucci, B. Wilson, C. A. Mattmann, and R. Chadwick, “New Capabilities for Prediction of High-Latitude Ionospheric Scintillation: A Novel Approach With Machine Learning,” Space Weather, vol. 16, no. 11, pp. 1817–1846, Nov. 2018, doi: 10.1029/2018SW002018.

[3] K. Lamb et al., “Prediction of GNSS Phase Scintillations: A Machine Learning Approach,” arXiv:1910.01570 [cs, stat], Oct. 2019, Accessed: Dec. 10, 2019. [Online]. Available: http://arxiv.org/abs/1910.01570.