loading page

Automated Bow Shock and Magnetopause Boundary Classification At Saturn Using Statistics of Magnetic Fields and Particle Flux
  • I Cheng,
  • Nick Achilleos,
  • Patrick Guio
I Cheng

Corresponding Author:[email protected]

Author Profile
Nick Achilleos
University College London
Author Profile
Patrick Guio
University College London
Author Profile


Statistical studies of the properties of different plasma regions, such as the magnetosheath and outer magnetosphere found near the boundaries of planetary magnetospheres, require knowledge of boundary (bow shock and magnetopause) crossings for purposes of classification. These are commonly detected by visual inspection of the magnetic field and / or particle data sampled by the relevant spacecraft. Automation of this type of activity would thus improve the efficiency of boundary and region studies, which benefit from large crossing datasets, and could also have implications for future development of onboard data-processing protocols in the pre-downlink stage. The Cassini mission at Saturn (2004-2017) provided an invaluable dataset for testing the viability of automated boundary classification. The training dataset consists of BS and MP crossings for the time period 2004 to 2016 (Jackman et al. (2019)). We have employed a series of techniques which involve pre-processing the calibrated magnetometer data, unsupervised training of a LSTM recurrent neural network on magnetometer data to filter magnetosheath regions where crossings are most likely to be found, isolating large rotations in magnetic field using minimum variance analysis (MVA), feature engineering such as magnetic field strength ratio either side of the field rotation to form a ‘feature vector’ for each candidate, and finally applying a gradient-boosting decision-tree-based algorithm to predict the probability that a given interval of data contains the signature of a bow shock (BS), a magnetopause (MP), or None (not a boundary crossing). The resulting model performs better on bow shock events, with a precision (fraction of true events in the retrieved sample) and recall (fraction of the total true events which were retrieved) of ~86% and ~90% respectively, as compared to ~50% and ~68% for the MP. The ongoing work focuses on augmenting the feature space for improved classification of MP, based on a magnetic pressure model of MP crossings derived using a local pressure balance condition (e.g. Pilkington et al. 2015) and using the distinct energetic particle flux changes across the MP in MIMI data (e.g. Liou et al. 2021). We expect that these promising new features will help us to better constrain the retrieval of candidate events which are true MP crossings.