1. SPACE WEATHER MISSION 02
Machine learning tools for
meteor shower characterization
in search of long-period comets
Andres Plata Stapper
Antonio Ordoñez
Jack Collison
Marcelo De Cicco
Susana Zoghbi
PLANETARY DEFENSE: MISSION 01
3. PLANETARY DEFENSE: LONG-PERIOD COMETS
Mission Statement
Provide more warning time for long period comet
impacts by applying machine learning to meteor
shower observations, whose trajectories enable
dedicated searches along predicted orbits.
4. PLANETARY DEFENSE: LONG-PERIOD COMETS
Objectives
● Improve and automate the identification of meteors on images detected in
meteor shower surveys using machine learning and deep learning
● Search for meteor shower streams and outbursts from the ever growing
meteor database to estimate parent body types and their associated
orbital parameters
● Find rare meteor outbursts from dust trail encounters that trace dangerous
long-period comets
14. PLANETARY DEFENSE: LONG-PERIOD COMETS
Describing the data
● Extract descriptive features, such as straightness
of trajectory, brightness of light curve, shape of
light curve, etc.
● Trained a random forest to classify meteor vs
non-meteor in dataset of ~200,000 objects from
CAMS
● Result: meteor classification precision = 90% and
recall = 81%
X
Y
Time
Intensity
15. PLANETARY DEFENSE: LONG-PERIOD COMETS
LSTM for time-series
● Long Short Term Memory (LSTM) networks can
efficiently characterize time-series
● Inputs: XY position, time, and intensity
● Result: meteor classification precision = 90%
and recall = 89%
Adapted from
http://karpathy.github.io/2015/05/21/rnn-effectiveness/
X1
Y1
T1
I1
X2
Y2
T2
I2
X3
Y3
T3
I3
H H H
σ
16. PLANETARY DEFENSE: LONG-PERIOD COMETS
Meteors vs. Non Meteors
Non Meteors Meteors
Clouds Planes Birds Small Bright Behind
Clouds
17. PLANETARY DEFENSE: LONG-PERIOD COMETS
Convolutional Neural Network (CNN)
Standard AlexNet architecture adapted to this dataset
Results: Precision: 88.6% Recall: 90.3%
23. PLANETARY DEFENSE: LONG-PERIOD COMETS
Multidimensional scaling via t-SNE
Dimension reduction
of meteor orbital
data via t-distributed
stochastic neighbor
embedding (t-SNE)
Meteor
25. PLANETARY DEFENSE: LONG-PERIOD COMETS
Established meteor showers
IAU meteor
shower
classification
correctly
rescued
by
t-SNE from
meteor orbital
parameters
Unlabeled
Meteors
26. PLANETARY DEFENSE: LONG-PERIOD COMETS
Unsupervised machine learning via DBscan
DBScan
Identifies the
established
showers in
addition to new
previously
undescribed
showers
28. Space time areas of
high density of
meteors
-
Outbursts
Allows for dedicated
searches for Long
Period Comets along
predicted orbits
Needs validation!
PLANETARY DEFENSE: LONG-PERIOD COMETS
Potential new outbursts
29. PLANETARY DEFENSE: LONG-PERIOD COMETS
Conclusions and breakthroughs
● Improved and automated the identification of meteors above human level
performance on images detected in meteor shower surveys using machine
learning and deep learning
● Recovered known meteor shower streams and characterized previously
unknown meteor showers from meteor orbital data
● We are able to find rare meteor outbursts from dust trail encounters that
could trace the orbits of dangerous long-period comets
30. PLANETARY DEFENSE: LONG-PERIOD COMETS
Acknowledgements
Frontier Development Lab
Peter Jenniskens
Pete Gural
Sara Jennings
James Parr
Siddha Ganju
JL Galache
Yarin Gal
FDL participants
NASA
Darlene Weidemann
Victoria Friedensen
SETI
Bill Diamond
IBM
Troy Hernandez
Graham Mackintosh
NVIDIA
Alison Lowndes
32. PLANETARY DEFENSE: LONG-PERIOD COMETS
Future directions
● Expand sky surveillance coverage
● Real-time monitoring of the meteor orbital data looking
for outbursts
● Publish search areas for the putative comets inferred
from the meteors
● Application that visually maps the comet orbit onto the
sky and provides search areas
34. Partition
historical
meteor orbital
data
Unsupervised
Machine
Learning
Validation
Showers and
Outbursts
From single meteor level orbital parameters to meteor showers and outburst characterization
Dimension
reduction
Total Number of meteor orbits 992226
Data partitioned in steps of 45° Solar Longitude
Data shown here corresponds to orbits 90-135° Solar Longitude
Total number of 122295
PLANETARY DEFENSE: LONG-PERIOD COMETS
Meteor shower detection