Regional TEC models developed by deep transfer learning

Caption: TEC coverage over the globe during March 2012 [see Habarulema, et al., 2018]. The aim of this project is to use transfer learning to adapt deep learning models developed for areas with good data coverage to work in areas with bad coverage.

Total electron content (TEC) is derived from fluctuations in GNSS signal and serves as an indication of ionospheric perturbations linked to solar activity. This is a useful metric for various space weather applications, such as uncertainty metrics in air craft positioning that are dependent on solar activity levels.

To build accurate prediction models for TEC large data sets are need wherever predictions are to be made. However, there are many regions with low data coverage, such as oceans, deserts, otherwise underdeveloped areas. The figure shows shows TEC variability across the globe. In this example regions along the geomagnetic equator in the south American sector has better data coverage than regions at similar latitudes in Africa. In general geomagnetic parameters have a strong latitudinal dependence, i.e. on average, areas at similar geomagnetic latitude behave similarly even if they are far apart in longitude.

The aim is to take advantage of the relatively dense data coverage in one region (say South America) and train a deep learning network to predict TEC over another region (e.g. central Africa). The first step would be to develop a deep-learning based TEC model to predict TEC in a region with adequate data coverage. Once this model shows satisfactory performance, it may be used as the basis for a TEC prediction model focussed on another region. By employing transfer learning the first model may be used to build an accurate model in a region with sparse data coverage.