spectral image based small target recognition...
TRANSCRIPT
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
SPECTRAL IMAGE BASED SMALL TARGET RECOGNITION
Stanley Grossman, Graduate Researcher
Dept. of Geography and Geoinformation Science
George Mason University
Fairfax, VA, USA
ABSTRACT
This paper will present the important factors influencing spectral target recognition using multispectral data and
suggest a different approach to small target recognition. The methodology of directed spectral search will be
offered, including the use of spectral libraries, spectral similarity scores, and ROC curve analysis. The ability to
find specific small targets, such as Saddam Hussein’s yellow taxicab typifies this challenge. This paper will
conclude with experimental results for a proposed technique of spectral image based small target recognition. It will
also suggest the nature of further research on the topic.
KEYWORDS: directed spectral search, multispectral, spectral similarity, small target recognition, NEF
INTRODUCTION
The nature of intelligence analysis has been changing rapidly since September 11, 2001. Prior to that date, the
primary focus for intelligence agencies had been on the traditional cold war enemies and their order of battle. High
spatial resolution panchromatic imagery and finely honed visual evidence-based analysis, known as literal
exploitation, was the preferred technique (Clark 2010). However, the current and future analytic focus has moved to
the war on terror and spotlighted the war of the individual. This has necessitated a change from exploitation based
on contrast and shape to one of signature and signal. Additionally, it moves the targets of interest from large-scale
in the tens of meters to ones of small-scale, typically a few meters and less. Moreover, the move to non-literal
exploitation (methods other than visual examination) also brings the requirement to exploit other regions of the
spectrum in addition to black and white pictures. Spectral data – color – becomes a significant discriminator in
analysis.
The “Yellow Taxicab Problem” is well known to most intelligence analysts and has come to epitomize a class
of target recognition problem – with a priori knowledge of what is to be identified, discover some unknown number,
if any, of those targets in an image. The moniker “Yellow Taxicab Problem” stems from the search for Saddam
Hussein in 2003. All the while Hussein was eluding Coalition pursuit, intelligence analysts knew that he was
frequently traveling around Baghdad in a yellow taxicab; however, analysts were unable to regularly and accurately
identify such a target in satellite imagery (Moore 2004). Similar analytic problems are common, such as the attempt
to determine the openness of elections in foreign countries. For instance, virtually all national security and police
forces of a certain country used similarly shaped and colored vans. Therefore, an attempt was made to identify vans
and determine from their spatial distribution and proximity to polling places if the security forces were unduly
influencing voting. In another classic case in Somalia, identifying “technicals” – militarized pick-up trucks usually
painted white (McConnell 2009) – was difficult to accomplish using space-based imagery. This concept can be
broadened to include other analytic protocols, such as finding specific freight containers on ship decks or order of
battle searches that might include mobile missile transporter erector launchers (Bergman 1996).
There are three primary reasons for these analytic failures. First, even with an influx of more technically savvy
analysts, generally, established analysts attempt to solve detection problems with literal methods by scrutinizing
high spatial resolution panchromatic imagery – using this method even when attempting to solve what should be
color-based queries. Second, there has been a historical dearth of truly high-resolution multispectral imagery able to
discriminate targets as small as an individual vehicle. Third, most target searches have used in-scene methods
starting with a seed from the scene and matching it to other possible scenic targets or to a material in spectral
databases (traditional image classification methodology). Seed pixels are not always identifiable or even available
in the scene nor valid from scene-to-scene. Adding complexity to the challenge, the physical size of critical targets
has shrunk. National Geospatial-Intelligence Agency (NGA) analysts have stated that upwards of 80% of all target
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
exploitation is for targets that are “small” (within the “blur” circle of the sensor) (Stenberg, Coleman, and Avilla
2003).
Until recently, commonly available multispectral imagery has been of relatively poor quality in relation to the
needs. The introduction of commercially available Geoeye and Worldview imagery is changing this situation
providing higher spectral and spatial resolution. Solving the yellow taxicab problem will entail using different
techniques as well. Therefore, this paper will examine the use of directed spectral search – starting with a seed
library signature modeled for the scene being analyzed – and recently available higher spectral and spatial resolution
multispectral imagery to attempt improved small target recognition.
BACKGROUND
The elemental premise in exploiting the visible spectrum is that the target can be characterized through the
nature of its reflectance (Schott 1997). Therefore, libraries of target spectra can be used to classify the target if the
image spectra have been calibrated to apparent (absolute) reflectance. The act of exploiting reflectance can be
broken down into three components (Gomez 2001):
• measurement of wavelengths
• measurement of intensities
• interpretation.
The measurement of wavelengths is accomplished by the construction of a sensor array of detectors for the
prescribed bands (e.g., detectors sensitive to photons in the 0.45-0.53µm wavelengths). The measurement of
intensities is more complex. On the surface it seems to be just a matter of counting the incident photons and
converting to reflectance; however, it is actually a complex process. As shown in Figure 1, photons incident on the
sensor may arrive from a number of sources other than the target pixel (Metzler 2010). In the figure, only path F is
solar illumination (insolation) that directly strikes the target and is reflected directly to the sensor. Every other path
is extraneous data from:
• background radiance (A)
• atmospheric radiance (upwelling radiance) (B)
• target emitted radiance (C)
• insolation scattered from the atmosphere (D)
• atmospheric emission reflected off the target (downwelling radiance) (E).
In order to properly identify a signature, these extraneous sources of energy must be accounted for and then properly
compensated for.
Figure 1. Potential photon paths to the sensor (Metzler 2010).
One interesting question related to spectral target recognition that persists is (Price 1994), “How unique are
spectral signatures?” The answer in the case of Dr. Price is a resounding “somewhat”. Historically, the association
A
B
C
D
E F
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
of a target to a unique signature has been hampered by several factors. The first was the relatively large pixel size of
the available imagery, such as Landsat 30m pixels. Additionally, traditional problems have included the
considerable spectral variability of samples (e.g., varying leaf colors within a plant species), confusion caused by
spectral similarity of two targets, and confusion caused by composite signatures (spectral signature mixing). New
high spatial resolution imagery available from Worldview-2 should address the large pixel issue with a nominal
pixel size for multispectral images of about 2m. Manolakis et al. (2003) broke the unique signature question into
four parts:
• “Does a spectrum uniquely specify a material”?
• “Are the spectra of any material the same when observed at different times”?
• “How is the spectrum of a ground material related to the spectrum observed by the sensor”?
• “How should we compare two spectra to decide if they are the same”?
METHODS
Study Data This study will analyze imagery acquired via Worldview-2 (WV-2). Launched in 2009, Worldview-2 offers
multispectral imagery with spatial resolution as high as 1.84 meters ground sample distance (GSD) at nadir and 2.08
meters at 20° off-nadir. WV-2 also provides high radiometric resolution with an 11-bit dynamic range. In addition
to the high spatial resolution, it offers increased spectral resolution with one PAN band and eight multispectral
bands spread across the visible and near infrared spectrum (DigitalGlobe, Inc. 2011). Available multispectral bands
are listed in Table 1. The study area is located in the El Segundo, CA vicinity (approximately 33° 54’ 52” N, 118°
23’ 38” W). It is representative of urban/suburban scenes. Study area J4-4010-p009-01 contains parking lots and
surface streets surrounding an electronics plant; scene J6-2050-p001-02 is a typical mixed-use neighborhood with
suburban streets and light business with parking lots; and scene J6-2050-p002-03 contains part of the airport tarmac,
parking lots, and feeder streets. The imagery was collected in June of 2011 on two different days when the climate
is Mediterranean-like with an average day-time temperature in the mid-70’s°F along with relatively clear and dry
days; spatial resolution in the study images was between 2.2m and 2.3m GSD.
Table 1. Worldview-2 multispectral bands and wavelengths.
Band Coastal Blue Green Yellow Red Red Edge
Near-IR 1
Near-IR 2
Wavelength (nm)
400-450 450-510 510-580 585-625 630-690 705-745 117-895 860-1040
Spectral Library The Nonconventional Exploitation Factors database (NEF) has been the National Geospatial-Intelligence
Agency’s (NGA) signature/material library program of record since the 1980’s. The NEF database includes
material properties that are laboratory measured to National Institute of Standards and Technology (NIST) traceable
standards detailing optical, thermal, bulk, and electromagnetic properties (National Geospatial-Intelligence Agency
2011). In addition to the directional hemispherical reflectance (DHR) information typical of signature databases, it
also includes a material taxonomy for bulk properties, surface properties, polarimetric properties, emittance
properties, bi-directional reflectance distribution function (BDRF) measurements, and radio frequency properties.
Spectral signature fidelity runs from 0.3 to 15µm at approximately 2nm increments, which provides band effective
data for multispectral and hyperspectral sensors. The most compelling capability of the NEF is the included
algorithms and models to calculate aperture radiance, aperture effective value, and laboratory effective value
spectral signatures that model a spectrum accounting for background, emissivity, atmosphere, collection geometry,
and illumination geometry parameters.
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
Directed Spectral Search Approach This study will assess target recognition against a database of spectral targets via directed spectral search. In
normal classification processes, several classes would be identified from within the scene and assessed against
training data, usually from the same image, and in situ “truth” (Congalton 1991). Exploitation via directed spectral
search starts with a suspected target or an a priori target – e.g., the yellow taxicab. Therefore, it will start with a
spectrum from a spectral database or library and necessitate a search of the image for one or more instances of the
target. The “yellow taxi cab” target recognition problem only requires two classes – the target and everything else.
Furthermore, the classification will be made in reverse - seeking within the scene for pixels with spectral signatures
matching a specific, tailored to the collection and solar geometry, signature from the database.
Figure 2. The research methodology includes creation of traditionally classified in-scene data for comparison to
results from the directed spectral search using a modeled library signature.
The data processing approach is performed in three parts: image preparation, signature generation, and target
analysis (Figure 2). In image preparation, the raw image data is transformed into a set of radiometrically
comparable signatures. After a suitable image scene has been selected, the image is masked to eliminate false
detections due to similarly colored background material that is obviously out of bounds (such as buildings), leaving
a subset image chip of roads, parking lots, and related surfaces where vehicles are likely to be found. A traditional
classification (see Accuracy Assessment) is performed to generate comparison target sets. Lastly, the image subset
must be corrected for atmospheric conditions. DigitalGlobe provides the calculations necessary to translate digital
counts in Worldview-2 image data to reflectance using calibration coefficients (Updike and Comp 2010).
Atmospheric correction is accomplished through three approximations: the cosines of the illumination and viewing
angles for atmospheric transmittance, dark pixel subtraction for upwelling radiance, and modeled data from the NEF
for downwelling radiance. Adding the atmospheric compensation components to the DigitalGlobe reflectance
formula produces:
(
)
( ) (1)
where ρ is the target reflectance, L is the target radiance, Lup
is the upwelling radiance, Ldown
is the downwelling
radiance, D is the distance to the sun, τs is the transmissivity along the solar path approximated by cos(θs), τv is the
transmissivity along the sensor viewing path approximated by cos(θv), Es is the band averaged solar irradiance
normal to the target, and the final cos(θs) represents the projected area effect of the solar angle to target. This
formula provides the top-of-the-atmosphere spectral reflectance signature for each potential in-scene target pixel.
The NEF signature library has been selected for the generation of search signatures because it offers
sophisticated algorithms to model spectral signatures for varied atmospheric conditions, collection geometries, and
solar geometries. In this case, the NEF Aperture Radiance (AR) algorithm was employed to generate search
signatures based on a modeled mid-latitude marine atmosphere suitable for the study site. AR utilizes scattered
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
insolation, scattered background radiance, upwelling atmospheric radiance, downwelling atmospheric radiance,
blackbody radiance, atmospheric transmission, material reflectivity, material emissivity and BDRF calculations to
generate a modeled signature as seen at the top of the atmosphere by the sensor. This signature is the effective
reflectivity (Metzler 2011) and simply described is the ratio of the aperture radiance for a material and the aperture
radiance of a reference material (spectralon) under identical conditions. Using WV-2 spectral response data, the
specified collection parameters from the imagery headers, and standard model atmospheres, AR generates
appropriately modeled spectral reflectance signatures for use as the a priori seeds.
Using Figure 1as a reference, the underlying radiometric AR calculation is (National Geospatial-Intelligence
Agency 2010):
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) (2)
where LA is the apparent radiance (total radiance reaching the sensor aperture from the direction of the material), τa
is the atmospheric transmission from material to sensor, ε is the directional emissivity, LBB is the blackbody radiance
at the target temperature T, ρA is the aperture effective directional reflectivity of the material’s surface, ↓LAE is
downwelling radiance due to atmospheric emission, ρs is the aperture effective bidirectional reflectivity in the solar
direction, LS is the effective exoatmospheric solar radiance, ↑LAE is the upwelling radiance due to atmospheric
emission, and λ is the wavelength.
Three spectra were chosen to represent the problem set and comprise a red, a blue, and a yellow material (Table
2). The red and blue signatures are paint over metal from widely available automobile models while the yellow is
paint on aluminum as a surrogate for an automobile.
Table 2. Description of the NEF spectral signatures used as seeds in the directed spectral search (National
Geospatial-Intelligence Agency 2011).
Description Spectral Description
Red, weathered paint on metal - 1990 GEO
Metro automobile
Reflectance rises rapidly from distinct minima near 0.35 and 0.50 microns to a peak reflectance near 0.735 microns. This peak is followed by a strong absorption band near 0.88 microns, and rapid fall off in reflectance toward 2.5 microns. The absorption bands near 0.35, 0.50 and 0.88 microns, identify the red pigment as ferric oxide, probably hematite.
Metallic-blue, weathered paint on metal - 1986 Toyota Corolla automobile
Strong pigment absorption bands near 0.33 and 0.60 microns (with a side band near 0.69 microns) result is a residual reflectance peak centered near 0.46 microns, imparting the blue color of this paint. The shape of the spectral curve is similar to that of ultramarine blue pigment. Reflectance rises steeply from 0.69 microns to 1.3 microns, after which it is relatively flat out to 2.5 microns.
Yellow, semi-gloss, zinc chromate primer
paint on aluminum
Overall reflectance rises very steeply in the visible to 0.73 microns, then more slowly to 1.32 microns, before declining toward 2.5 microns. Absorption by an unknown pigment near 0.36, 0.60 (shoulder), and 0.83 microns account for the yellow color. For BRDF, the difference in the ss and sp bistatic scans at 0.6328 and 1.05mm indicates that the sample is primarily a volume scatterer. The sample is glossy and becomes specular as the wavelength increases.
The target analysis phase performs the directed spectral search and accuracy analysis. Even though the
classification starts at the library and reverses the classification direction, traditional spectral matching algorithms
may be employed. Any of the popular algorithms are potentially applicable. However, for this research the spectral
similarity score (SSS) algorithm was chosen. The spectral similarity score combines the results of spectral angle
mapper (SAM) with a Euclidean distance (ED) calculation to create a hybrid value (Nidamanuri and Zbell 2010).
SAM converts a known and an unknown spectrum each to a set of vectors in n-vector space where n is the number
of bands and then compares the angular distance between the two vectors. A threshold is used to determine when
two signatures are in close enough proximity to be labeled as similar with smaller angles corresponding to better
matches. When used with calibrated imagery, SAM is not influenced by topographic illumination, geometry
variations, or general albedo differences because it ignores vector magnitude and only thresholds on angular
distance (C. Hecker et al. 2008; Aspinall, Marcus, and Boardman 2002; Kutser, Miller, and Jupp 2006; The
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
University of Texas at Austin, Center for Space Research). The Euclidean distance is a measure of the difference in
magnitude of the two vectors varying between zero and one (Nidamanuri and Zbell 2010). Euclidean distance
potentially improves the method’s ability to find targets in shadows (Coogan 2011). The combined SAM and ED
result is sensitive to both the shape and magnitude of each spectral signature (Nidamanuri and Zbell 2011). The SSS
is built with formulas 3, 4, and 5 (C. Hecker et al. 2008; Luc et al.).
(∑
√(∑
)(∑
)
) (3)
where n is the number of spectral bands, t is the spectral reflectance of the target signature and r is the spectral
reflectance of the known signature.
√
∑ ( )
(4)
where s1 and s2 are the target and known spectra respectively and n is the number of bands.
√( ) (5)
Accuracy Assessment In typical classification processes accuracy assessment is accomplished through the presentation of some form
of confusion matrix with counts and percentages of pixels correctly and incorrectly identified. However, the
intelligence analyst’s goal is not number of pixels but number of targets and the determination of success or failure
at the fine granularity of the pixel level would most likely lead to the possibly erroneous conclusion of poor
recognition performance. The proper determination of target recognition is whether the target is positively
highlighted; stipulation by one pixel or a dozen pixels is not relevant. The very nature of the test is that of a binary
predictor and the use of receiver operating characteristic (ROC) curves may be a better measure of realization
(Gönen). ROC curves assist in assessing the accuracy of binary predictions. The key component of ROC curves is
the comparison of rates, for instance, the rate of positively predicting a result versus negatively predicting a result
(Barnes et al. 2007). These rates are presented as a related curve of sensitivity against specificity. In the case of this
study, there are three illustrative ROC sensitivities: 1) probability of detection (Pd) representing the number of true
predictions divided by the sum of true targets; 2) probability of false detections (Pfd) which is the number of false
predictions over the sum of the false predictions and true targets; and 3) probability of missed detections (Pmd) being
the number of true targets not predicted over the sum of true targets. The ROC curve is then produced by
juxtaposing a rate (sensitivity) against the SSS threshold (specificity). A point is generated for each predictive
threshold execution of the directed spectral search (e.g., target predictions with an SSS value less than or equal to
0.15). This study incremented thresholds by 0.025 from a no-detect value to an obvious over-detect value for each
scene and signature. These data pairs (e.g., [Pd, SSS] or [Pfd, SSS]) are then fit to a curve through cubic
interpolation. The intersection of the Pfd curve with the Pmd curve represents an analytic “sweet spot” where any
change in Pd will trade false detections for missed detections. This sweet spot is an arbitrary contrivance for
accuracy assessment. Actual intelligence problems would dictate threshold trade-off; for instance, the need to
absolutely find as many targets as possible may compel accepting higher false detection rates but fewer missed
targets while a need to only correctly identify a target may compel lower false prediction rates while increasing the
risk of missing valid targets. Comparison of ROC curves provides a relative method for judging acceptable
performance.
Because ground truth was not available for these images (as is typically the case for intelligence analysts), a
different performance benchmark was required. The ENVI® target detection wizard (ITT Visual Information
Solutions 2010) was engaged using the mixture tuned matched filter (MTMF) method with minimum noise fraction
(MNF) transform and standard ENVI® defaults to generate sets of traditionally classified in-scene targets.
Atmospheric compensation was performed using the ENVI® dark object subtraction method. The results of these
classifications were used in lieu of ground truth and served as performance benchmarks for comparison to the
experimental results.
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
RESULTS
The execution of this experiment resulted in three sets of data with each set consisting of the directed spectral
search against three signatures. This produced nine sets of ROC curves, one for each signature for each scene,
presented in Figure 3 through Figure 5. In each set, there are four curves: a red curve for performance solely against
the MTMF generated list of targets labeled “constrained” and a set of blue curves for performance against all targets
in the scene (including those not discovered by MTMF) labeled ”total”. The solid curves are the probability of
detection (Pd) against the constrained targets (red) or against all targets (blue); the dashed curve is the probability of
missed targets (Pmd); and the dot-dashed curve is the probability of falsely detected targets (Pfd). The “analytic sweet
spot” is the point where the Pmd and the Pfd cross in each set of curves. This is the point at which any change in Pd
requires trading missing detections for false detections. The sweet spot is identified with an arrow and three
statistics, 1) the spectral search score associated with the sweet spot, 2) the total Pd associated with the point, and
3) the Pfd associated with the point. The red arrow with a single statistic is the Pd performance solely against
constrained targets at the associated total (blue) sweet spot.
J4_4010_p009_01 Scene J4_4010_p009_01 included city streets, parking lots, and airport tarmac of both concrete and light
asphalt. In this scene, searching for vehicles with the modeled red signature performed well (Figure 3a). Against all
targets, the directed spectral search found 90% of the constrained targets and 86% of the total targets at the sweet
spot. The Pfd remained low at 13% for the total targets. The directed spectral search for the yellow vehicles (Figure
3b) at the sweet spot recognized 74% of all targets against a 26% Pfd; however, 95% of the constrained targets were
identified at this point. Against blue targets (Figure 3c), the algorithm again performed well with a total Pd of 82%
and 86% of the constrained targets identified.
J6-2050-p002-03 A typical El Segundo street scene is depicted in J6-2050-p002-03 (Figure 4) with suburban streets, houses, and
office and retail parking lots. Against this backdrop, red vehicles (a) were consistently identified with a total Pd of
81% and a constrained targets recognition rate of 85% against a 19% Pfd. Yellow vehicles (b) were harder to detect,
succeeding only 64% of the time but finding 75% of constrained targets. Blue vehicle detection (c) was in the
middle with a total Pd of 69% and a constrained success of 84%.
Figure 3. ROC curves for recognition rates in scene J4_4010_p009_01. The blue curves correspond to performance
rates against all targets in the scene (total performance) while the red curve represents performance against just the
MTMF generated targets.
b) a) c)
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
J6-2050-p002-02 Scene J6-2050-p002-02 (Figure 5) contains the immediate environs of a typical light industry (electronics) plant
with parking lots and a few city streets. Yet again, the red vehicle (a) detection performance is highest with a total
detection rate of 90% and a constrained detection rate of 94% with only 11% Pfd. Detection rates on yellow vehicles
(b) are anomalously low at only 40% total and 43% of constrained targets. Blue detection rates (c) once more are
moderately successful at 71% total Pd and 79% detection rate for constrained targets.
In terms of raw numbers (Table 3), the directed search performed significantly more robustly than the mixture
tuned match filter method, usually by a factor of two or more. It is clear that the MTMF search left a substantial
number of targets undiscovered and that the directed search covered this gap quite well. It must be noted that that
while in many cases the probability of detection is lower against the total set of targets than the constrained set, this
probability score represents a hefty increase in the actual number of targets detected. For instance, in the case of the
red targets in scene J4_4010_p009_01, the directed search discovered 90% of the MTMF generated targets but in
summation found 2¼ times the number of MTMF targets. The MTMF search only found 38% of the actual number
of targets in the scene (549).
Figure 4. ROC curves for recognition rates in scene J6-2050-p002-03. The blue curves correspond to performance rates
against all targets in the scene (total performance) while the red curve represents performance against just the
MTMF generated targets.
Figure 5. ROC curves for recognition rates in scene J6-2050-p002-02. The blue curves correspond to performance rates
against all targets in the scene (total performance) while the red curve represents performance against just the
MTMF generated targets.
b) a) c)
b) a) c)
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
Table 3. Comparison of the number of targets found at the sweet spot versus available targets. Base is the number of MTMF
generated targets discovered; additional is the number of targets found above the base by the directed search (SSS),
found is the sum of base and additional; and total is the number of targets generated (MTMF) and actually in the scene
(SSS). The base and found ratios are SSS:MTMF while the total ratio s MTMF:SSS.
ANALYSIS
With one exception, detection performance was similar from scene to scene for a specific signature search. For
instance, for the red signature in all scenes, directed spectral search performed relatively well, finding on average
85% of the total targets and 90% of the constrained targets at the analytic sweet spot. For the blue signatures, it
performed moderately well, finding on average 83% of the constrained targets and 74% of all targets. The yellow
signature directed spectral search performed feebly with an average of 59% against all targets but 71% of the
constrained targets at the sweet spot. However, there are anomalously low results for the yellow signature in scene
J6-2050-p002-02. This is most likely due to the statistically small number of true yellow targets (four constrained
targets and seven total targets). Removing these anomalous data yields somewhat better results of 69% total Pd and
85% constrained Pd. Directed spectral searches for red signatures performed better in all scenes while yellow (with
exclusion of the anomalous data) and blue directed spectral searches performed similarly in the moderate range.
There did appear to be a scene-based bias with searches in scene J4_4010_p009_01 outperforming the other two
scenes.
The yellow signature results are somewhat suspect, especially those for J6-2050-p002-02. In all scenes for the
red and blue vehicles, there were adequate sample sizes from near 100 to several hundred total targets. In scenes
J4_4010_p009_01 and J6-2050-p002-03, there were at least 30 yellow target vehicles allowing for reasonable
results. However, in scene J6-2050-p002-02, there were only seven total yellow targets imbuing each target
counted, missed, or falsely chosen to be overly weighted. When the detection threshold reaches a sufficient
distance, the red and yellow targets are confused so that at some point falsely detected red targets overwhelm truly
detected yellow targets. In J4_4010_p009_01 and J6-2050-p002-03, this phenomenon occurs well after the majority
of the yellow targets are detected and only a few remain. However, in scene J6-2050-p002-02, because of the low
number of valid targets, when 38% of the targets are identified (only three targets) there are already eleven false
detections, a Pfd of138%.
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
Figure 6. Normalized histogram of target detections versus spectral similarity scores for J4_4010_p009_01.
Normalized histograms of the final data run for each chip and signature are presented in Figure 6 through
Figure 8. The frequency of total target detection pixels (true and false) are binned by their respective spectral
similarity scores. As expected, red and yellow signature plots show a left (negative) skew with more target
pixels detected as the SSS threshold increases. While the blue signature histograms are left skewed as well,
they also present a bi-modal distribution. Examination of the detection rates in the blue signature data shows
that as the SSS threshold increases, the number of falsely detected targets eventually exceeds the number of
truly detected targets. Ultimately, this situation reverses and the second modal presentation occurs. This is
likely caused by addition of false pixels due to shadows mimicking blue targets. Eventually, the vehicle
signature similarity passes beyond the shadow signature; the curve drops back to a typical skew pattern picking
up additional true targets causing the bi-modal nature of the histogram.
Figure 7. Normalized histogram of target detections versus spectral similarity scores for J6-2050-p002-03.
The results demonstrate the influence of collection geometry on reflectance. Both J6-2050-p002-02 and J6-
2050-p002-03 are chips from the same image while J4_4010_p009_01 was collected two days earlier. Because
WV-2 is sun-synchronous, the solar geometries are comparable in both scenes. However, while the sensor elevation
is similar in the two images, the sensor azimuth is considerably different at 51.4° for J4_4010_p009_01 and 283.1°
for J6-2050-p002-02 and -03. The difference in geometry places the shadows in direct line-of-sight during imaging
for J4_4010_p009_01 but not for the other chips. This causes the more sharply defined bi-modal distribution for the
blue signature histogram and a general right-shift in all the histograms for J4_4010_p009_01.
Figure 8. Normalized histogram of target detections versus spectral similarity scores for J6-2050-p002-03.
b) a) c)
b) a) c)
b) a) c)
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
The statistical mean for each of the target detection histograms correlates well with the related spectral
similarity score at the analytic sweet spot; therefore, the histogram mean would be a suggested starting SSS
threshold for analysis with actual thresholds tuned up or down depending on the nature of the analysis. The
standard deviations range from 0.02 through 0.07 with eight out of nine being between 0.02 and 0.04; only the
J4_4010_p009_01 blue histogram is outside that range.
Table 4. The strong correlation between the target detection mean and sweet spot spectral similarity score
provides an analytic starting point for exploitation.
Red Yellow Blue
Mean SSS Mean SSS Mean SSS
J4_4010_p009_01 0.46 0.49 0.38 0.38 0.35 0.36
J6-2050-p002-03 0.34 0.33 0.30 0.31 0.34 0.38
J6-2050-p002-02 0.34 0.35 0.28 0.28 0.32 0.35
SUMMARY AND RECOMMENDATIONS
Spectral small target recognition was demonstrated using a directed spectral search technique. Preliminary
results showed that directed spectral search is a viable alternative for intelligence analysis and suggested a metric to
compare results (analytic sweet spot) as well as an algorithmic threshold for initial analysis. The studied method did
compensate for atmosphere, collection geometry, and illumination geometry; however, the results demonstrated
expected variability based on target-to-background contrast and collection geometry. Future work will compare
additional algorithms, the effect of better background characterization specific to an individual target, the
contribution of spectral unmixing on performance, and the influence of collection geometry on results.
REFERENCES
Aspinall, R. J., W. A. Marcus, and J. W. Boardman. 2002. “Considerations in collecting, processing, and analysing
high spatial resolution hyperspectral data for environmental investigations.” Journal of Geographical
Systems 4 (1) (March 25): 15-29.
Barnes, L. R., E. C. Gruntfest, M. H. Hayden, D. M. Schultz, and C. Benight. 2007. “False Alarms and Close Calls:
A Conceptual Model of Warning Accuracy.” Weather and Forecasting 22 (5) (October): 1140-1147.
doi:10.1175/WAF1031.1.
Bergman, S. M. 1996. The Utility of Hyperspectral Data to Detect and Discriminate Actual and Decoy Target
Vehicles. Master of Science in Systems Technology (Scientific and Technical Intelligence), Monteray, CA:
Naval Postgraduate School, December.
C. Hecker, M. van der Meijde, H. van der Werff, and F. D. van der Meer. 2008. “Assessing the Influence of
Reference Spectra on Synthetic SAM Classification Results.” IEEE Transactions on Geoscience and
Remote Sensing 46 (12) (December): 4162-4172. doi:10.1109/TGRS.2008.2001035.
Clark, R. M. 2010. Intelligence Analysis. 3rd ed. Washington, DC: CQ Press.
Congalton, R. G. 1991. “A review of assessing the accuracy of classifications of remotely sensed data.” Remote
Sensing of Environment 37 (1): 35–46.
Coogan, J. 2011. Conversations between Deputy Chief for Advanced Geospatial Services and Chief Scientist,
Physical Sensing Technology Center, Raytheon IIS.
DigitalGlobe, Inc. 2011. DigitalGlobe: Worldview-2 Satellite. DigitalGlobe.
http://www.digitalglobe.com/index.php/88/WorldView-2.
Gomez, R. B. 2001. Spectral Library Issues In Hyperspectral Imaging Applications. In Proc. 5th Joint Conf.
Standoff Detection for Chemical and Biological Defense. Williamsburg, VA, September 24.
Gönen, M. Receiver Operating Characteristic (ROC) Curves. Statistics and Data Analysis. SAS Institute, Inc.
ITT Visual Information Solutions. 2010. ENVI 4.8.
Kutser, T., I. Miller, and D. L.B Jupp. 2006. “Mapping coral reef benthic substrates using hyperspectral space-borne
images and spectral libraries.” Estuarine, Coastal and Shelf Science 70 (3): 449–460.
ASPRS 2012 Annual Conference
Sacramento, California ♦ March 19-23, 2012
Luc, B., B. Deronde, P. Kempeneers, W. Debruyn, and S. Provoost. “Optimized Spectral Angle Mapper
classification of spatially heterogeneous dynamic dune vegetation, a case study along the Belgian
coastline.” Image 1: 2.
Manolakis, D., D. Marden, and G. A Shaw. 2003. “Hyperspectral image processing for automatic target detection
applications.” Lincoln Laboratory Journal 14 (1): 79–116.
McConnell, T. 2009. Terrorist Saleh Ali Saleh Nabhan killed by US commandos in Somalia. September 16.
http://www.timesonline.co.uk/tol/news/world/africa/article6836129.ece.
Metzler, M. 2011. NEFDS Exploitation Algorithms presented at the NEFDS Training at NGA Campus East, August
31, Springfield, VA.
Metzler, M. 2010. Overview/Surface Optical Material Properties presented at the NEFDS Training at White Sands
National Test Center, April 13.
Moore, R. 2004. Hunting Down Saddam: the Inside Story of the Search and Capture. New York: St. Martin’s Press.
National Geospatial-Intelligence Agency. 2010. Nonconventional Exploitation Factors Data System Modeling
Document. 12.2 ed. March 15.
National Geospatial-Intelligence Agency. 2011. NEF Data System Documentation.
Nidamanuri, R. R., and B. Zbell. 2010. “A method for selecting optimal spectral resolution and comparison metric
for material mapping by spectral library search.” Progress in Physical Geography 34 (1) (February): 47-58.
doi:10.1177/0309133309356376.
Nidamanuri, R. R., and B. Zbell. 2011. “Normalized Spectral Similarity Score (NS3) as an Efficient Spectral Library
Searching Method for Hyperspectral Image Classification.” IEEE Journal of Selected Topics in Applied
Earth Observations and Remote Sensing 4 (1) (March): 226-240. doi:10.1109/JSTARS.2010.2086435.
Price, J. C. 1994. “How unique are spectral signatures?” Remote Sensing of Environment 49 (3): 181–186.
Schott, J. R. 1997. Remote Sensing - The Image Chain Approach. New York: Oxford University Press.
Stenberg, G, J Coleman, and F Avilla. 2003. Conversations between Deputy Chief for Advanced Geospatial
Services and NGA Senior Scientists (PL).
The University of Texas at Austin, Center for Space Research. Spectral Angle Mapper Classification. Analysis of
Hyperspectral Imagery. http://www.csr.utexas.edu/projects/rs/hrs/analysis.html.
Updike, T, and C Comp. 2010. Radiometric Use of WorldView-2 Imagery | Technical Note. DigitalGlobe, Inc.,
November 11.