Title: Exemplar Spectra
1EXEMPLAR SPECTRA
2Exemplar Spectra Topics
- Exemplar Spectra Definition
- Exemplar Spectra Advantages
- Exemplar Spectra Applications
- Exemplar Spectra Generation
- Exemplar Spectra Case Study
3Exemplar Spectra Definition
Individual Spectra
Exemplar Spectrum
Spectral Analysis Synthesis
Exemplars are single spectra, derived from a set
of individual measurements, that capture the
spectral essence of a class or subclass of
materials
4Exemplar Spectra AdvantagesBetter Data Quality
Spectra to be included in a library often cover a
wide range of instruments, measurement protocols,
sample characteristics, collection environments,
etc.
- Even under highly controlled conditions, it is
standard scientific procedure to make multiple
measurements and average them together to reduce
noise
Typically results in substantial variability in
spectral response not associated with fundamental
material characteristics
5Field Measurement Variability Example 1
- Four spectral scans of one painted panel show
clear variability - Difference between days likely due to inaccuracy
in Lambertian assumption - Other differences due to noise, environment, or
measurement protocol
Enlargement of SWIR region
3/9 44 3/13 10 3/9 45 3/13 11
6Exemplar Spectra Advantages Data Quantity
Reducing data volume should improve performance
by allowing focus on exploitation rather than
data management and by expanding range of classes
considered
- Difficult for either analysts or algorithms to
effectively sort through and use 1000s of
spectra, particularly in directed search
application
Notional Numbers
500 spectra ? 100 classes or 10,000 spectra ?
2,000 classes
10,000 spectra ? 100 classes
7Exemplar Spectra Advantages Spectral Expertise
- Using large set of individual spectra requires
analyst to evaluate and decide which is more
reliable, which features are noise and which
are information, which features are generally
diagnostic - in short, requires more spectral
expertise on part of every analyst - Exemplar generation by spectral experts should
allow broader pool of analysts to effectively
process and exploit hyperspectral data
8Exemplar Spectra Advantages
- A spectral library populated by exemplars should
provide more robust representations of class
spectral signatures - Removing noise, artifacts of measurement process,
etc. - Emphasizing class-level diagnostic spectral
features vs. sample-specific spectral features - A spectral library populated by exemplars should
support more efficient exploitation - Reducing the total number of samples to be
considered - Focusing attention on diagnostic features
- Allowing expansion of range of classes considered
- A spectral library populated by exemplars should
allow a broader pool of less experienced analysts
to effectively exploit hyperspectral data - Incorporate spectral expertise at exemplar
generation stage
9Exemplar Spectra Applications
- Exemplars are intended for use in spectral
matching - - Finding something in an image that matches a
training sample - - Target detection or material identification
are spectral matching applications - - Target characterization is NOT a spectral
matching application
10Exemplar Spectra Applications Detection /
Identification
Is it a smiley face?
- Compare features of unknown sample to features of
training sample - Two eyes
- Overall round shape
- Yellow color
- Curved mouth with curves on ends
- Degree to which unknown matches training sample
establishes confidence of detection /
identification
Training
Unknown
Key Element Looking for some known and
distinctive entity - distinguishing between
distinct classes / materials
11Exemplar Spectra Applications Characterization
How happy is it?
Estimate condition of known sample (e.g., soil
moisture, snow age, chemical concentration) Spect
ral matching to training set is impractical -
discrete cases of continuous variable
Training
Unknown
Key Element Estimating a continuous variable for
a sample whose identity has been determined
12Spectral LibraryCharacteristics vs. Intended Uses
13Exemplar Spectra Generation
Compile all spectra in Spectral Library
associated with class of interest
Sample Selection
Spectra
Evaluate quality of spectra, select final set for
processing
Sample Review
Remove channels strongly affected by atmospheric
absorption, select common channel set
Channel Subsetting
Exemplar Spectra
Identify, separate spectrally distinct groups
within label class
Spectral Clustering
Derive weighted least squares best fit to each
spectrally distinct group
Local Regression
Fill descriptor fields (including exemplar
quality)
Metadata Generation
14Step 1 Sample Selection
- Use NSEC Spectral Library search / extract
capability to produce a single composite file or
a list of individual files - Example Select all samples identified as Eastern
Bloc paints - Result List of candidate spectra
15Step 2 Sample Review
- Look at each spectrum, evaluate utility for
exemplar generation - presence of severe artifacts, noise
- large gaps in wavelength range
- extraneous sources of variability (e.g., angular
orientation) - Remove spectra from set as necessary
- Result Revised list of spectra to be used in
exemplar generation
16Step 3 Atmospheric Filter
Based on atmospheric transmission curves and
obvious atmospheric artifacts in spectra, define
exclusion windows Result Consistent set of
wavelength samples for all spectra that avoids
major atmospheric effects
Transmission based windows
Adjusted windows
17Step 4 Spectral Clustering
Apply spectral-angle-based clustering
algorithm Review results, split, combine, or
delete clusters as appropriate Result One or
more sets of spectrally similar clusters 1
cluster 1 exemplar
18Step 5 Local Regression
- Apply local regression (LOESS) to cluster members
- Response at each wavelength estimated using a
quadratic fit to the data points near the
wavelength of interest - Fit determined using weighted least squares (more
weight to samples closer to wavelength of
interest) - Nearness defined by user specified smoothing
parameter - Tends to remove noise spikes without changing
broader features (in contrast to other smoothing
techniques)
Original Spectra
LOESS fit
19Exemplar Spectra Example 2
- Field measurements from earlier example - same
panel, two days - LOESS fit produces reasonable representation of
curve shape, removes noise artifacts
Enlargement of SWIR region
Exemplar 3/9 44 3/13 10 3/9 45 3/13 11
20Step 6 Produce Metadata
ltSpectral Library Entry Typegt, Exemplar
Spectrum ltMetadata Chapter Versiongt,
1.3 ltSecurity Classificationgt, unclassified ltEntry
Name gt, CARC 383 Green Paint ltData Quality gt,
high ltConstituent Spectra gt, t309rcg.044,
t309rcg.045, . . ltIndependent Variable Type gt,
wavelength ltIndependent Variable Units gt,
micrometers ltDependent Variable Type gt,
reflectance factor . . . ltend_metadatagt 0.40,
0.065334 0.405, 0.064188 . . . 2.4, 0.366493
Interactively fill required and appropriate
metadata fields for library Result NSEC library
compatible ASCII file containing exemplar and
associated metadata
21Exemplar Trade Study
- Goal Provide a quantitative assessment of the
advantages or disadvantages of using exemplars in
target detection/ID applications - Approach
- Select data set with reasonable amount of target
and background diversity (HYDICE Forest Radiance
II) - Develop exemplars representing target classes in
test set, using archival spectra - Compensate for atmospheric effects (MODH2O)
- Apply detection/identification algorithms (NIDA,
CEM) using as training - Exemplars
- All exemplar source spectra
- Ground measurements of actual targets
- Score performance, comparing training
alternatives - All processing to be done at the National Air
Intelligence Center (NAIC)
22Forest Radiance II
- Aberdeen Proving Grounds, Maryland - June 2000
- Realistic deployments of U.S. and foreign
vehicles, nets
23Exemplar Spectra to be Generated
- U.S. CARC 383 Green Paint
- SLISE panels (Field)
- Western Rainbow (Field)
- JHU (Lab)
- AFRL (Lab)
- Forest Radiance I (Field)
- Desert Radiance I (Field)
- Eastern bloc Green Paint
- Southern Rainbow (Field)
- Desert Radiance II (Field)
- Forest Radiance I (Field)
- U.S. LCSS Woodland Net
- Southern Rainbow (Field)
- JHU (Lab)
- Desert Radiance I (Field)
- Misc. (Lab)
- East German Net
- Southern Rainbow (Field)
- SITAC (Lab)
- Swedish Barracuda Net
- no sources found
24Constrained Energy Minimization (CEM)
- Description
- Spectral matched filter with normalizing value in
denominator - perfect match to target produces
value of 1 - NAIC version uses variant of covariance matrix,
includes mechanism to iteratively purify
background distribution (recompute excluding
pixels with highest detection statistic values)
25NAIC Identification Algorithm (NIDA)
- Combines Constrained Energy Minimization (CEM),
Spectral Angle (SA), and BANDS - CEM applied with low threshold to detect most
targets (typically with many false alarms) - SA computed between CEM detections and training
spectra, relaxed threshold applied to remove
least target-like detections (presumably false
alarms) - BANDS applied to remaining detections to identify
targets - Uses high order derivatives of 6th order
polynomial fit to spectra to ID local maxima and
minima - Compares maxima and minima to interactively
derived training signature - Produces binary decision (target / non-target)
for each training spectrum
Applied Analysis, Inc. (Huguenin and Jones,
1986)
26Trade Study Part 1 - CEM Scoring
- Produce ROC curves for each scene, and for all
scenes combined, from detection statistic images
generated using - Exemplar spectra
- Each individual spectrum
- Compare performance (scene-by-scene and
composite) - percent detected at fixed false alarm rate
- false alarm rate at fixed detection percentage
- Compare scene-to-scene performance variability as
a function of training approach - Best high detected, low false alarm rate,
low variability
27Trade Study Part 1 - NIDA Scoring
- Compare performance (scene-by-scene and
composite) for each target type and all target
types combined - Percent correctly identified at fixed false alarm
rate - False alarm rate at fixed level of correct
identification - Confusion matrices
- Compare scene-to-scene performance variability as
a function of training approach - Best high correct ID, low false alarm rate,
low variability