Application of Machine Vision Technology to Martian Geology Ruye WangHarvey Mudd College James Dohm University of Arizona Rebecca CastanoJet Propulsion.

Application of Machine Vision

Technology to Martian Geology

Ruye Wang Harvey Mudd CollegeJames Dohm University of ArizonaRebecca Castano Jet Propulsion Laboratory

AISRP 2004-2007

April 4, 2005

ObjectivesObjectives

►Develop an intelligent system for Develop an intelligent system for robust detection and accurate robust detection and accurate classification in multispectral classification in multispectral remote sensing image dataremote sensing image data

►Demonstrate system in context of Demonstrate system in context of Martian geology applicationMartian geology application

ApproachApproach► Pre-conditioningPre-conditioning

Modified PCAModified PCA Decorrelation StretchDecorrelation Stretch Conversion to emissivityConversion to emissivity

► UnsupervisedUnsupervised Kohonen competitive networks Kohonen competitive networks K-MeansK-Means

• Euclidean DistanceEuclidean Distance• Spectral Angular Mapping (SAM)Spectral Angular Mapping (SAM)

Independent Component Analysis (ICA)Independent Component Analysis (ICA)► SuperviseSupervise

Support Vector Machine (SVM)Support Vector Machine (SVM) Other statistical and neural network methodsOther statistical and neural network methods

Application to Martian Application to Martian geologygeology

►Two regions selected for focused study Thaumasia highlands Coprates Rise mountain range

Study MotivationStudy Motivation► The Wind River Range and the two Martian mountain

ranges display similar features such as magnetic signatures, thrust faults, complex rift systems, and cuestas and hogbacks.

Field-based mapping indicates that the Wind River Range records a Late Archean history of plutonism that extends for more than 250 m.y. The range is dominated by granitic plutons, including gneiss, batholith, and granites.

Martian mountain ranges are ancient based on their magnetic signatures. What about their compositions?

• The detection of mountain-building rocks would provide critical clues to the evolution of the core, mantle, and crust on Mars.

Study Objectives/RationaleStudy Objectives/Rationale► Use new tools to investigate the hypothesized

diversity of rocks and minerals in the selected regions Compare to previously reported compositions Identify materials of low abundance that previous

techniques may not have been sensitive enough to identify

► Compare the selected Martian regions to the Wind River Mountain Range in Wyoming Identify if the mountain ranges under investigation

contain mountain-building rock materials such as metamorphic and silicic-rich plutonic rocks as identified in the Wind River Mountain Range

Comparison of Martian and proposed Comparison of Martian and proposed Earth Analog SitesEarth Analog Sites

Coprates Rise mountain range Mars

Wind River Mountain Range Wyoming

Cuestas and hogbacks, which are caused by tectonic tilting and differential erosion, are visible at both sites

Martian Multispectral DataMartian Multispectral Data► THermal EMission Imaging System (THEMIS)

on Mars 2001 Odyssey orbiter spacecraft Low spectral resolution (multi-spectral):

►10 IR channels (6.78-14.88 microns)►5 VIS channels

High spatial resolution:►IR 100m/pix, VIS 18m/pix

► Thermal Emission Spectrometer (TES) on Mars Global Surveyor spacecraft High spectral resolution (hyper-spectral):

►143 or 286 channels (6.25-50 microns) Low spatial resolution:

►3000m/pix

Site Selection Site Selection

Themis Image ofThaumasia Highlands

USGS Geological Map(based on Viking image)

K-Means Clustering with SAM Distance

► Shades/shadows in rugged mountain areas do not reflect spectral properties;

► Use spectral angles mapping distance (SAM):

1cosu v

u vα − ⎡ ⎤⋅= ⎢ ⎥

⎣ ⎦

Competitive Learning Clustering with Normalized Vectors

►Normalize both weight and data vectors to consider angular difference only

Clustering Results

► K-means (SAM, Euclidean)

► Competitive net (Kohonen)

Comparison of Spectral Angular Map

and Euclidean Distance

Modified PCA and Decorrelated Stretch

Modified PCA and Decorrelated Stretch

Original Themis image

First three PCAs

Decorrelated Stretch

Support Vector Machine Support Vector Machine (SVM)(SVM)

►Linear separationLinear separation

0, ( ( )) 1,( )

0, ( ( )) 1,T y sign f X X P

f X X W by sign f X X N

> = = ∈⎧= + =⎨< = =− ∈⎩

1

m

j j jj

W y Xα=

=∑

Support Vector Machine Support Vector Machine (cont)(cont)

Non-linear separation by kernel mapping( )X Xφ→

( ) ( )1

( ) ,m

T

j j jj

f X X W b y K X X bφ α=

= + = +∑

( ) ( ) ( )1 2 1 2,T

K X X X Xφ φ=

Support Vector Machine Support Vector Machine DemoDemo

Support Vector Machine Support Vector Machine ExampleExample

From left to right: Training Results Themis image Context

Future WorkFuture Work

► Exploration of TES dataExploration of TES data► Conversion from radian data to Conversion from radian data to

emissivity emissivity ► Application of independent component Application of independent component

analysis (ICA)analysis (ICA)► Usage of spectral library data for Usage of spectral library data for

supervised trainingsupervised training

Comparison between Comparison between THEMIS and TESTHEMIS and TES

► THermal EMission Imaging System (THEMIS) Low spectral resolution (multi-spectral):

►10 IR channels (6.78-14.88 microns)►5 VIS channels

High spatial resolution:►IR 100m/pix, VIS 18m/pix

► Thermal Emission Spectrometer (TES) High spectral resolution (hyper-spectral):

►143 or 286 channels (6.25-50 microns) Low spatial resolution:

►3000m/pix

Independent Component Analysis Independent Component Analysis (ICA)(ICA)

► In low-spatial resolution image, the spectrum of a pixel may be linear mixture of multiple end-members.

If spectra of end-members are known, least squares methods are used to separate them. [M. Ramsey et al 1998]

Otherwise this is a blind source separation problem, which may be addressed by ICA algorithms.

( ) ( ), 1, ,j jj

x s nλ α λ λ= =∑ K

Independent Component Analysis Independent Component Analysis (cont)(cont)

► Given m linear mixtures (pixels) of n end-members:

► Estimate abundances and spectral signatures for the end-members.

€

Independent Component Analysis Independent Component Analysis (cont)(cont)

► Given Given m linear mixtures (pixels) of n end-members:

► Estimate abundances and end-member spectral signatures

1W A−=

s Wx=

x As=or

Obtain Temperature and Emissivity

from Radian data

[e.g., A. Gillespie et al 1998, S. Liang, 2001]

Backup SlidesBackup Slides

Thaumasia highlands and Coprates rise mountain ranges record magnetic signatures, thrust faults, complex rift systems, and cuestas and hogbacks [Dohm et al., 2001], possibly indicative of a plate tectonic phase during extremely ancient Mars [Baker et al., 2002]

3-D oblique view using MOLA data looking to the west across Valles Marineris (C) and the Thaumasia plateau (white ine). Also shown are the locations of the Thaumasia highlands (A) and Coprates rise (B) mountain ranges with respect to Valles Marineris (C), Syria Planum (D), and Tharsis Montes (E). The mountain ranges are ancient as observed in the MGS-based magnet data (Acuna et al., 1999) and structural mapping of Dohm et al., 2001a).

TR

RS

ST

WR

T

R

CF

T

200 km

Left. MOLA topographic map showing the west-central part of the Thaumasia highlands mountain range, which includes thrust faults (T), complex rift systems (R), shield volcanoes (s), fault systems such as Claritas Fossae (CF), and locales such as Warrego Rise (WR) interpreted to be centers of magmatic-driven uplift and associated volcanism, tectonism, and hydrothermal activity (Anderson et al., 2001). Warrego Rise forms the highest reach within the mountain range. Right. 3-D topographic projection merged with layers of paleotectonic and paleoerosional information of the Warrego Valles source region (Dohm et al., 2001)

Detailed geologic map of the northeast part of theThaumasia region (Dohm et al., 2001). Geologic map units(colored polygons), faults (yellow lines), and ridges (black lines)are shown.