Automated detection and unsupervised classification of electron enhancement events in Earth’s magnetotail
Presentation Type
Talk
Presenter Format
In Person Meeting Talk
Topic
Nightside Science
Start Date
11-5-2022 11:30 AM
Abstract
We present a layered algorithm approach for the detection and unsupervised classification of both dispersed and dispersionless electron enhancement events utilizing MMS data from the tail seasons of 2017 through 2021. Using a Savitzky-Golay filter over cumulative relativistic electron count rates from the Fly’s Eye Energetic Particle Spectrometer (FEEPS) (80 – 150 keV), we find approximately 40,000 events over 5 seasons of data, greatly expanding the spatial and temporal coverage (extending up to 30 RE in Earth’s magnetotail) of known injection events. A portion of the events detected in the 2018 tail season was compared with a manual, SME-generated, list of events and we find that the algorithm is able to correctly identify 96% of the labeled dispersionless injections in addition to finding an additional ~5k events over the temporal extent of the list. Spatial patterns in the distribution of algorithm-identified events were binned according to driving solar wind conditions and geomagnetic indices to uncover potential physical drivers, and additionally binned by location in the tail region. We find the spatial distribution of identified events correlates well with previously identified injection mechanisms such as central plasma sheet crossings, bursty bulk flows (BBFs), and magnetotail reconnection, while perhaps finding a new driving mechanism near the magnetopause flanks. In addition to FEEPS measurements, data from MMS’ fluxgate magnetometer (FGM) and Fast Plasma Investigation (FPI) were utilized for unsupervised characterization of each event, utilizing either a finite time period bordering each identified enhancement or the full extent of an event from onset until cumulative electron counts returned to near pre-enhancement levels. We explore utilizing a variety of data dimensionality techniques for multivariate time series (MTS) data sets, such as PCA similarity metrics and the extended Frobenius norm (Eros) distance measure to quantify similarity across different events while accounting for time-based covariance of features in MTS data. Distance measures were then used to compute distinct clusters of events in an unsupervised setting using algorithms such as k-Nearest Neighbors (kNN) and t-distributed Stochastic Neighbor Embedding (t-SNE). Representative examples from the resulting clusters were then analyzed to assess the ability of the unsupervised learning to both identify and differentiate between known physical processes, using known example cases from the MMS data set. Known example cases spanned physical phenomena including: i) central plasma sheet (boundary) crossings, ii) bursty bulk flows, iii) magnetotail reconnection, and iv) phenomena adjacent to the magnetopause flanks.
Automated detection and unsupervised classification of electron enhancement events in Earth’s magnetotail
We present a layered algorithm approach for the detection and unsupervised classification of both dispersed and dispersionless electron enhancement events utilizing MMS data from the tail seasons of 2017 through 2021. Using a Savitzky-Golay filter over cumulative relativistic electron count rates from the Fly’s Eye Energetic Particle Spectrometer (FEEPS) (80 – 150 keV), we find approximately 40,000 events over 5 seasons of data, greatly expanding the spatial and temporal coverage (extending up to 30 RE in Earth’s magnetotail) of known injection events. A portion of the events detected in the 2018 tail season was compared with a manual, SME-generated, list of events and we find that the algorithm is able to correctly identify 96% of the labeled dispersionless injections in addition to finding an additional ~5k events over the temporal extent of the list. Spatial patterns in the distribution of algorithm-identified events were binned according to driving solar wind conditions and geomagnetic indices to uncover potential physical drivers, and additionally binned by location in the tail region. We find the spatial distribution of identified events correlates well with previously identified injection mechanisms such as central plasma sheet crossings, bursty bulk flows (BBFs), and magnetotail reconnection, while perhaps finding a new driving mechanism near the magnetopause flanks. In addition to FEEPS measurements, data from MMS’ fluxgate magnetometer (FGM) and Fast Plasma Investigation (FPI) were utilized for unsupervised characterization of each event, utilizing either a finite time period bordering each identified enhancement or the full extent of an event from onset until cumulative electron counts returned to near pre-enhancement levels. We explore utilizing a variety of data dimensionality techniques for multivariate time series (MTS) data sets, such as PCA similarity metrics and the extended Frobenius norm (Eros) distance measure to quantify similarity across different events while accounting for time-based covariance of features in MTS data. Distance measures were then used to compute distinct clusters of events in an unsupervised setting using algorithms such as k-Nearest Neighbors (kNN) and t-distributed Stochastic Neighbor Embedding (t-SNE). Representative examples from the resulting clusters were then analyzed to assess the ability of the unsupervised learning to both identify and differentiate between known physical processes, using known example cases from the MMS data set. Known example cases spanned physical phenomena including: i) central plasma sheet (boundary) crossings, ii) bursty bulk flows, iii) magnetotail reconnection, and iv) phenomena adjacent to the magnetopause flanks.
