TY - JOUR
T1 - Classification of Cassini’s Orbit Regions as Magnetosphere, Magnetosheath, and Solar Wind via Machine Learning
AU - Yeakel, Kiley L.
AU - Vandegriff, Jon D.
AU - Garton, Tadhg M.
AU - Jackman, Caitriona M.
AU - Clark, George
AU - Vines, Sarah K.
AU - Smith, Andrew W.
AU - Kollmann, Peter
N1 - Funding Information:
KY, JV, GC, SV, and PK were supported by an JHU APL internal research and development grant, funded by the Civil Space Mission Area within the Space Exploration Sector. KY and JV were additionally supported by NASA Cassini/MIMI mission contract NNN06AA01C. GC was additionally supported by NASA’s Cassini Data Analysis program under grant 80NSSC18K1234, which also partially supported KY and SV was additionally supported by NASA grant 80NSSC19K0899. TG’s work at the University of Southampton was supported by the Alan Turing Institute and the Science and Technology Facilities Council (STFC). CJ’s work at DIAS was supported by the Science Foundation Ireland (SFI) Grant 18/FRL/6199. AS was supported by STFC Consolidated Grant ST/S000240/1 and NERC grants NE/P017150/1 and NE/V002724/1.
PY - 2022/5/20
Y1 - 2022/5/20
N2 - Several machine learning algorithms and feature subsets from a variety of particle and magnetic field instruments on-board the Cassini spacecraft were explored for their utility in classifying orbit segments as magnetosphere, magnetosheath or solar wind. Using a list of manually detected magnetopause and bow shock crossings from mission scientists, random forest (RF), support vector machine (SVM), logistic regression (LR) and recurrent neural network long short-term memory (RNN LSTM) classification algorithms were trained and tested. A detailed error analysis revealed a RNN LSTM model provided the best overall performance with a 93.1% accuracy on the unseen test set and MCC score of 0.88 when utilizing 60 min of magnetometer data (|B|, Bθ, Bϕ and BR) to predict the region at the final time step. RF models using a combination of magnetometer and particle data, spanning H+, He+, He++ and electrons at a single time step, provided a nearly equivalent performance with a test set accuracy of 91.4% and MCC score of 0.84. Derived boundary crossings from each model’s region predictions revealed that the RNN model was able to successfully detect 82.1% of labeled magnetopause crossings and 91.2% of labeled bow shock crossings, while the RF model using magnetometer and particle data detected 82.4 and 74.3%, respectively.
AB - Several machine learning algorithms and feature subsets from a variety of particle and magnetic field instruments on-board the Cassini spacecraft were explored for their utility in classifying orbit segments as magnetosphere, magnetosheath or solar wind. Using a list of manually detected magnetopause and bow shock crossings from mission scientists, random forest (RF), support vector machine (SVM), logistic regression (LR) and recurrent neural network long short-term memory (RNN LSTM) classification algorithms were trained and tested. A detailed error analysis revealed a RNN LSTM model provided the best overall performance with a 93.1% accuracy on the unseen test set and MCC score of 0.88 when utilizing 60 min of magnetometer data (|B|, Bθ, Bϕ and BR) to predict the region at the final time step. RF models using a combination of magnetometer and particle data, spanning H+, He+, He++ and electrons at a single time step, provided a nearly equivalent performance with a test set accuracy of 91.4% and MCC score of 0.84. Derived boundary crossings from each model’s region predictions revealed that the RNN model was able to successfully detect 82.1% of labeled magnetopause crossings and 91.2% of labeled bow shock crossings, while the RF model using magnetometer and particle data detected 82.4 and 74.3%, respectively.
KW - boundary crossings
KW - Cassini-Huygens
KW - machine learning
KW - magnetosphere
KW - random forest
KW - recurrent neural network (RNN) long short-term memory (LSTM)
KW - Saturn
UR - http://www.scopus.com/inward/record.url?scp=85131753734&partnerID=8YFLogxK
U2 - 10.3389/fspas.2022.875985
DO - 10.3389/fspas.2022.875985
M3 - Article
AN - SCOPUS:85131753734
SN - 2296-987X
VL - 9
JO - Frontiers in Astronomy and Space Sciences
JF - Frontiers in Astronomy and Space Sciences
M1 - 875985
ER -