Autonomous Mineral Detectors for Mars Rovers and Landers

1
Autonomous Mineral Detectors for Mars Rovers and
Landers

• Martha S. Gilmore1, Rebecca Castaño2, Ben
  Bornstein2, Jim Greenwood1,
• Matt Merrill1

1 Wesleyan University, Middletown,CT 2 Jet
Propulsion Lab, Pasadena, CA
AISRP PI Meeting April 4, 2005
2
Objectives
Goal
Design and develop software to enable rovers to
autonomously analyze spectral data and identify
data indicating geologically important signatures.
Non-carbonates
Carbonates
Spectral data
3
Approach

• Develop detectors for minerals of interest
• Select mineral set
• Detector design
• Detector training
• Detector testing
• Lab data
• Field data
• Orbital data
• Issues
• Dust coating
• Detector sensitivity
• Non-linear mixing
• Develop novelty detector

4
Mineral Class Selection

• Minerals associated with water, hydrothermal
  minerals
• Carbonates (calcite), sulfates (jarosite),
  evaporites, phyllosilicates, phosphates, iron
  oxides
• Common rock forming minerals
• Other silicates (e.g., pyroxene, olivine)
• Dust - Mars Soil Simulant (palagonite)

5
Detector Design

• Supervised classification
• Neural networks
• Support Vector Machines
• Binary detectors mineral class present or not
• Preprocessing of data
• Band selection
• Data are reflectance (comparison to white
  standard)
• Key is to obtain sufficient training data
• Training data are mixtures of library spectra

6
Training

• Synthetic rocks are a linear mixture of JPL,
  ASTER, USGS, and other end-member spectra
  (n1600).
• Mixture percentages and their likelihoods are
  drawn from uniform random distributions bounded
  by ranges we specify.
• We have adjusted the rocks to be consistent with
  martian petrology including the martian
  meteorites.

7
Generative Training Data Model
r(m, b)
wm
Mixture Model
?m wm (r(m, b) gb,m(0,s))
Synthetic Rocks
8
Carbonate Detector

• Feedforward neural network trained using
  backpropogation
• Network architecture determined empirically
• Modularized for portability
• Uses 2000-2400 nm range
• Source code and weight file lt 100 KB
• Initial testing on 32 field data measurements
  were 100 successful (8 carbonates, 24
  noncarbonates) (Gilmore et al, 2000)

9
Dust Coatings Experiment
Aliquots of dust are blown into the air and
settle onto a calcite crystal.
The spectra are taken and dust thickness measured
under a microscope.
10
Dust Coated Spectra
11
Carbonate detector

• Activation value proxy for band depth (BD) a
  2300 nm
• Threshold change in rate of decrease of BD
  values and BD 0.1
• Appropriate for calcite veins on martian surface

12
Jarosite Detector
Kernel
Mineral found at Opportunity outcrops
13
(No Transcript)
14
Gypsum CaSO4.2H2O and jarosite on Earth
Jarosite and a dissolved blade-like mineral on
Mars
15
Chemical Analysis
Jarosite verified by X-ray diffraction, SEM -
Energy Dispersive Spectrometry and Spectral
Reflectance in the lab
16
Results
17
Non-linear Mixing

• Produced mixtures of carbonate and dust, basalt
• Quantify the sensitivity of the carbonate
  detector
• Issues with smallest (lt44 mm) grain sizes
• Examine the effects of intimate mixtures
• Spectra measured under controlled lab conditions
  to model non-linear behavior - modeling in
  progress.

18
Novelty Detection
2s
1s

• Novelty detection
• Unsupervised clustering to detect outliers

19
Infusion Plan

• Integrate onto CLARAty testbed
• Demonstrating our detectors onboard the FIDO
  research rover in an autonomous science traverse
  scenario (funded by the Onboard Autonomous
  Science Investigation System (OASIS))
• Method baselined for AFL mission by JPL
• Run detectors on AVIRIS, OMEGA, CRISM data

20
Benefit

• Increase mission science return
• Point-spectra are cheap and this analysis has low
  computing cost
• Long-term, long range roving, this software can
  be used to analyze spectra taken during a
  traverse (no measurements otherwise scheduled)
• Data products

• Compiled statistics can be returned
• Identification of a mineral of interest and data
  sample for the rock that would otherwise have
  been overlooked
• Science alert

21
Backup Slide
22
Minerals
23
Spectral Database