OnBoard Processing for Spectral Remote Sensing

1
On-Board Processing for Spectral Remote Sensing

• Richard B. Gomez
• Ambrose J. Lewis
• Center for Earth Observing and Space Research
• School of Computational Sciences
• George Mason University
• Fairfax, Virginia

ISPRS Special Session Future Intelligent Earth
Observing Satellites (FIEOS) 10 -15 November
2002, Denver, Colorado
2
Agenda

• Hyperspectral Imaging Overview
• Hyperspectral Imaging Introduction
• Hyperspectral Cube
• Hyperspectral and Multispectral
• Real-Time Systems Overview
• Hardware Clustering Techniques
• Spectral Library Overview
• Spectral Library Introduction
• Spectral Library Functional Requirement
• Exemplar Spectra
• Conclusions

3
Hyperspectral Introduction

• Hyperspectral Remote Sensing is a very powerful
  capability!!!
• It is possible to determine composition of
  materials from afar
• Hundreds of available spectral bands provide
  much more data
• With Hyperspectral Remote Sensing come unique
  challenges
• Dataset size becomes enormous
• Users need to visualize in a 200 dimensional
  space
• Must account for Hughes Phenomenon
  Atmospheric Effects
• Hyperspectral Applications include
• Biological Chemical Detection
• Homeland Security
• Disaster Mitigation and Environmental
  Monitoring
• Traffic Flow
• Law Enforcement

4
Reflected and Emitted Energy
UV
BLUE
GREEN
RED
NIR
SWIR
MWIR
LWIR
What you see is not what you get!
5
Hyperspectral Cube
6
Hyperspectral Cube

Pixel Spectrum
Flight Line
Intensity
Single Pixel
Wavelength
Spatial Pixels
Spectral Bands
Single Sensor Frame
Series of Sensor Frames
7
Hyperspectral Multispectral
Multispectral
Hyperspectral
each pixel has several large discrete
spectral bands
each pixel has many small continuous spectral
bands
R
R
l
l
8
Data Space Representations

• Spectral Signatures - Physical Basis for Response

• N-Dimensional Space - For Use in Pattern Analysis

9
Real-Time Systems

• Real-Time Systems (RTS) are utilized in
  applications that required deterministic
  performance
• Temporal aspects of system operation can be
  guaranteed
• Example applications include navigation, control,
  medical, and sensors
• RTS can have a minimal Quality of Service
  determined to meet system functional requirements
• RTS are suitable for mission critical
  environments
• fault tolerant
• high-availability

10
Hardware Clusters

• It may be feasible to increase the on-board
  processing power of remote sensing satellites by
  developing clusters of platform-based computers
• Utilized in parallel to efficiently process the
  remotely sensed data before transmission to the
  ground
• The various processors could look for different
  solutions within the same dataset (e.g.
  dedicating a processor to a search for a
  particular type of data)
• The clustering technique has been exploited with
  tremendous success on Earth-based problems
• A generic class of clustered computers developed
  by NASA is the so-called Beowulf systems
• Beowulf clusters leverage low to no cost software
  (MPI), operating systems (Linux), and hardware

11
PC/104 Linux Mini-clusters

• A promising area for the space environment is
  work being done at by the Embedded Reasoning
  Institute (ERI) of Sandia National Laboratory
• PC/104 is an IEEE standard (IEEE P996.1) that
  describes a single-board computer (SBC)
• Given the fairly small size and low power
  consumption of PC/104 computers, they may be
  suitable for the satellite environment
• http//eri.ca.sandia.gov/eri/howto.html
• These computers are compliant with industry
  standardized hardware and software of the PC
  architecture
• A significant benefit of this approach is the
  availability of robust software and operating
  systems

12
Example PC/104 CPUs
www.parvus.com
www.rtdusa.com
13
Spectral Library Introduction

• Spectral Libraries provide reference spectra to
  compare against remotely sensed data
• Pure pixel end members are compared against
  library spectra in N-dimensional space
• Search determine likely spectral matches
• Rest of the pixels are treated as combinations
  of multiple materials, pixel unmixing
  algorithms are used to decompose into component
  parts
• Each spectra in the library contains a few
  hundred points and appropriate metadata
• Spectra Library should have taxonomy to
  facilitate analyst navigation

2-D Endmembers
200-D Endmembers
?
band i
band j
14
Spectral Library Functional Requirements
Spectral Library
Atmospheric Corrected HSI Data
Spectra 1
metadata repository
R
l
Spectra 2
comparison operator
data match
R
l

taxonomy based search
metadata match
Spectra N
R
l
Analyst Query
15
Objectives

• Design a spectral library architecture that can
• Perform fast and efficient spectral matching
• Allow processing at a level of resolution
  appropriate for each classification task
• Store library reference spectra without loss of
  information

16
Spectral Library Parameters

• 1. Size of the spectra to be stored
• 2. Number of spectra to be stored
• 3. Similarity between stored spectra
• 4. Precision of stored data
• 5. Associated metadata
• 6. Efficient library search
• 7. Scalability and modifiability of the library
• 8. Multiple data formats and encoding schemes

17
Optimizations to be Performed

• Find the correct library spectrum by performing a
  minimum number of pair-wise spectral comparisons
• Perform the simplest calculations necessary for
  quick and accurate spectral comparison
• Minimize the number of features of our unknown
  spectrum needed for comparison

18
Spectral Comparison

• Rather than comparing the unknown spectrum with
  every reference spectrum in the library, we only
  need to descend the tree structure to the group
  containing the correct reference spectrum (i.e.,
  the best match)
• This requires that we have a way to compare the
  unknown signal with entire groups of spectra,
  rather than the particular spectra contained in
  each group

19
Library Class Tree
20
Exemplar Spectra
Derived from a set of individual measurements,
exemplar spectra are single spectra that capture
the spectral essence of a class or subclass of
materials
21
Conclusions

• The information technology (IT) data processing
  approach shows promise
• This IT approach offers a method to process and
  exploit, on board, huge amounts of data without
  the spectral analyst in the loop
• Use of exemplar spectra and XML methodology in
  HSI applications will facilitate real time
  spectral analyses
• Development of standards would help

22
Conclusions

• Hierarchical library structure for quick,
  accurate and efficient searching has been
  developed
• Spectra is clustered into hierarchical groups of
  high and low separability
• The process minimizes total number of
  comparisons
• Complex spectral comparisons are done only where
  it is necessary

23
Closing Remarks and QA