Presentation #108.06 in the session Improving Understanding of the Sun-Earth System Through Advanced Statistical and Machine Learning Techniques.
In the post-sunset ionosphere rising plumes of low-density plasma known as equatorial plasma bubbles (EPBs) can form. EPBs can degrade GNSS signals and, communications between ground stations and spacecraft. As our dependence on space-based assets continues to grow, there is an ever-pressing need to better understanding and ultimately forecast EPB events.
SWARM is a 3-satellite LEO mission that launched into a near polar orbit in 2013 at an altitude of 460 and 540km. We first create a labelled dataset of EPB’s using a Savitzky-Golay smoothing filter. Comparing it to the existing EPB-labeller on-board SWARM we see the classification accuracy improves from 75% to 81%.
We then feed the labelled set into a random forest classifier (RFC) using longitude, spacecraft potential, plasma density and ion temperature as features. The former is the most important feature identified by the RF and the longitudinal dependence of EPBs has long been reported. Potential was the second most pertinent feature and the link between spacecraft charging and EPBs does not appear to have been previously reported. Spacecraft charging is primarily the product of electron temperature and density. Therefore, it can be viewed as a naturally occurring engineering feature for the model, thus explaining its importance in the RF. The model recall is 95% which is a respectable performance as our objective is to detect EPBs and therefore to minimize false negatives. The F1 score is 88% and we also deem this acceptable. Future work will extend the study into forecasting EPBs and integrating additional data.