SHARE 2012 - RIT Center for Imaging Science

Transcription

SHARE 2012 - RIT Center for Imaging Science
SHARE 2012: Large Edge Targets for Hyperspectral Imaging
Applications
Kelly Canham, Daniel Goldberg, John Kerekes, Nina Raqueno, and David Messinger
Center for Imaging Science
Rochester Institute of Technology
54 Lomb Memorial Dr.
Rochester, NY, 14623 USA
ABSTRACT
Spectral unmixing is a type of hyperspectral imagery (HSI) sub-pixel analysis where the constituent spectra and abundances
within the pixel are identified. However, validating the results obtained from spectral unmixing is very difficult due to a lack
of real-world data and ground-truth information associated with these real-world images. Real HSI data is preferred for
validating spectral unmixing, but when there is no HSI truth-data available, then validation of spectral unmixing algorithms
relies on user-defined synthetic images which can be generated to exploit the benefits (or hide the flaws) in the new
unmixing approaches. Here we introduce a new dataset (SHARE 2012: large edge targets) for the validation of spectral
unmixing algorithms. The SHARE 2012 large edge targets are uniform 9m by 9m square regions of a single material (grass,
sand, black felt, or white TyVek). The spectral profile and the GPS of the corners of the materials were recorded so that the
heading of the edge separating any two materials can be determined from the imagery. An estimate for the abundance of two
neighboring materials along a common edge can be calculated geometrically by identifying the edge which spans multiple
pixels. These geometrically calculated abundances can then be used as validation of spectral unmixing algorithms. The
size, shape, and spectral profiles of these targets also make them useful for radiometric calibration, atmospheric adjacency
effects, and sensor MTF calculations. The imagery and ground-truth information are presented here.
Keywords: Hyperspectral, ground-truth, and SHARE 2012
1. INTRODUCTION
Ground-truth information is needed to validate many image processing techniques and to calibrate aerial and space-based
sensors. Acquiring ground-truth information is an extensive task in both time and money. During September of 2012, a
large ground-truthing campaign was conducted at Avon, NY, by the Digital Imaging and Remote Sesning (DIRS) Laboratory (see Figure 1 for an overview of the Avon site). During this campaign targets for spectral unmixing, radiometric
calibration, unique signatures, and several other tests were deployed.1 The campaign lasted one day and was imaged by
over four different aerial sensors and two space-based sensors. Two of the many targets deployed during this experiment
were a set of large (9m square) bright and dark tarps that neighbored grass and sand background materials. These targets
can be used for many different purposes including radiometric calibration, spectral unmixing, atmopsheric adjacency analysis, and MTF calculations. Here we describe these large targets and possible uses for them.
Ground-truth material spectral libraries have been generated for a handful of sites around the world. Two of the most
well-known images that have extensive ground-truth and/or spectral libraries are Cuprite, Nevada, and Indian Pines, Indiana.2–5 These two datasets are featured extensively as a means of improving or validating new algorithms. This is because
these two sites have freely available imagery and ground-truth information. The data were imaged by the HSI AVIRIS sensor and the ground-truth data were collected on the day of collection. SHARE 2012 data can be included in this short-list
of validation imagery once the data have been made available to the scientific community.
Further author information: (Send correspondence to Kelly Canham or David Messinger.)
Kelly Canham: E-mail: [email protected]
David Messinger: E-mail: [email protected], Telephone: +1 585 475 4538
Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XIX,
edited by Sylvia S. Shen, Paul E. Lewis, Proc. of SPIE Vol. 8743, 87430G · © 2013 SPIE
CCC code: 0277-786X/13/$18 · doi: 10.1117/12.2016271
Proc. of SPIE Vol. 8743 87430G-1
Downloaded From: http://spiedigitallibrary.org/ on 08/07/2013 Terms of Use: http://spiedl.org/terms
Figure 1. Wide-Area Sensor Platform (WASP) image of the Avon, NY, SHARE 2012 site. Zoomed in to show the black and white large
targets and their location to other SHARE 2012 targets and materials.
This paper is separated such that Section 2 discusses the targets themselves, Section 3 discusses the natural materials
surrounding the targets, Section 4 briefly describes the some of the applications which can use the targets and associated
imagery, and Section 5 discusses how these targets will be used in future work within the DIRS Laboratory.
2. SHARE 2012 LARGE TARGETS
During the SHARE 2012 field campaign three large targets (or tarps) were deployed; one black felt, one white TyVek®, and
one gray felt. The uniformity of the black target was measured prior to SHARE 2012 by using an ASD spectroradiometer
with the bare-fiber held approximately 1m above the target so that the area measured covered a 1m area (including a portion
of the operators legs and feet). Seven transects were completed on the black felt target; each transect was spaced approximately 1.5m from the previous transect and a total of 10 spectral files were saved and each file consists of 10 averaged
spectra (for a total of 100 individual spectral measurements during each transect). Averaging all of the relative spectral
reflectance from this uniformity analysis gives the average spectral reflectance profile of the target as shown in Figure 3.
Additionally, a covariance plot of the spectral information was compiled and is presented in Figure 2(b). This covariance
matrix calculates the spectral covariance between each of the 70 (seven transects with 10 spectral files) reflectance measurements to all the others collected over the entirety of the tarp. A spectral angle metric (SAM6 ) was also computed
between all the spectra, and the resulting SAM matrix is shown in Figure 2(c). The black diagonal line in Figure 2(c)
represents an exact spectral match (or a SAM of 0.0) and white reprents a SAM value of 0.248. This uniformity analysis
was collected on a clear Rochester day with the tarp laid out on top of a gray aphalt parking lot and may be subject to
minor atmospheric conditions. A 4in square Spectralon white reference was used to optimize the ASD and convert the
input radiance measurements into relative reflectance. Similar uniformity measurements will be conducted for the white
and gray tarps as future work (dependent on clear Rochester weather conditions).
All three large targets deployed during SHARE 2012 measured 9m by 9m (with an error of no more than 10cm deviation from the 9m requirement). The size was determined so that each target was approximately nine pixels along an
edge for a 1m GSD sensor (such as a panchromatic image from Quickbird or IKONOS or from the SpecTIR HSI aerial
sensor flown at an altitude of approximately 2500ft). The size was selected to guarentee at least one pure material pixel
at the center of each target while at the same time containing enough edge pixels to be used in a novel spectral unmixing
approach.7 The material of each target was selected to be spectrally flat and to span the dynamic range of most remote
sensing sensors. The ground-truth data from SHARE 2012 shows the black felt has an average reflectance value of 7%,
the white TyVek has an average reflectance value of 81%, and the gray tarp has an average reflectance of 26% over the
Proc. of SPIE Vol. 8743 87430G-2
Downloaded From: http://spiedigitallibrary.org/ on 08/07/2013 Terms of Use: http://spiedl.org/terms
Average Black Tarp Spectrum
0.1
160
0.09
Spectral Covariance Histrogram of Large
Black Tar. et
L
0.08
140
N
0.07
'-''
0.06
0.05
0.04
0.03
u
7
Ñu
O2
N
Q
O >
O
0.02
0.01
Z
120
80
60
40
20
0
400
900
1400
1900
T
.II.'Fnh
100
MPPriiLL
o
0.0005
2400
0.0006
Wavelength [nm]
0.0007
0.0008 0.0009
Covariance Value
(a)
0.001
0.0011
0.0012
(b)
SAMetric Histogram
400
Ñ3
Q3
1n
L 250
Ñ 200
ái
150
1ñ
. 100
50
2
0
0.05
0.1
0.15
0.2
0.25
SAMetric Value
(c)
Figure 2. Black target uniformity analysis. Figure 2(a) shows the average black tarp spectrum, Figure 2(b) shows the black tarp covariance matrix, and Figure 2(c) shows the black tarp SAM matrix.
visible through SWIR wavelengths (400-2500nm). Figure 3 shows the SHARE 2012 average spectral profiles for each of
the targets (black, gray, and white). The average spectral profiles are a result of averaging 100 individual spectra from
the same material together and were collected for an approximate 6cm diameter spot size (using a 3°fore-optic positioned
approximately 1m above the ground/target).
The layout of the three large targets during SHARE 2012 was carefully considered so that multiple analyses could
be performed using the same targets. The gray tarp was positioned in a parking lot that was covered in old gravel and
asphalt. The black and white targets were positioned side-by-side to form a highly contrasting straight edge between them.
It was determined that the large contrasting targets should be placed in a location such that multiple background materials
would form an edge with the targets. As such, a sand volleyball pit surrounded by a grass field within the Avon park was
determined to be the ideal location for the black and white target. The placement of these two targets allowed for two edges
of the black target to be adjacent to grass, one edge of the white tarp to be adjacent to the grass, one edge of the black tarp
to be adjacent to sand, and two edges of the white tarp to be adjacent to sand (as seen in Figure 1).
The edge between the black and white targets was positioned so that it falls at a heading of approximately 11° from
the originally planned SpecTIR sensor’s flight-line. This angle is needed for both the novel spectral unmixing validation
approach as well as for MTF calculations, but deviations from this angle occurred due to wind conditions during overflight.
Due to the wind’s impact on the actual flight-line, measured GPS coordinates of the targets are more important than originally planned. The GPS coordinates for the eight corners (two of which are duplicated along the black-white edge) are
presented in Table 1. From this information and the imagery, validation of the orthorectication process can be completed
as well as determining the exact angle of the black-white edge in the imagery.
Proc. of SPIE Vol. 8743 87430G-3
Downloaded From: http://spiedigitallibrary.org/ on 08/07/2013 Terms of Use: http://spiedl.org/terms
Large Target Spectral Profiles
Black Tarp
White Tarp
Gray Tarp
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
400
900
1900
1400
2400
Wavelength [nmj
Figure 3. Average spectral profile of the three large targets deployed during SHARE 2012.
Table 1. Black and White Targets Corner GPS Coordinates
Corner
Northeast
Southeast
Southwest
Northwest
Black Tarp
42.90783°N, -77.7652°W
42.90775°N, -77.7653°W
42.90776°N, -77.7654°W
42.90784°N, -77.7654°W
White Tarp
42.90781°N, -77.7651°W
42.90773°N, -77.7652°W
42.90775°N, -77.7653°W
42.90783°N, -77.7652°W
Table 2.
3. BACKGROUND MATERIALS
The placement of the black and white large tarps in relation to the background materials was determined so that the different material edges can be used for both a novel spectral unmixing validation approach as well as a means to quantify
the atmospheric adjacency effect and MTF calculations. During SHARE 2012 field work, the reflectance of each of these
background materials was measured. The same number of spectra were collected for the background materials as were
collected for the targets (100 spectra) using the same collection configuration (6cm spot size). During the spectral collection process large swaths of these background material areas were walked so that spatial variability could be considered
within the different materials.
The sand proved to be a good medium upon which to stake the targets because lawn stakes were easy to fasten and
remove and the sand could be raked to form a flat surface upon which to layout the targets. During a test deployment, it was
determined that footprints and water content within the sand (due to recent rain and dew) caused undesirable BRDF affects,
therefore a large area (approximately 10m by 20m), south of both the black and white targets, was raked prior to imaging
and reflectance measurements. By raking the sand a more uniform area with consistent BRDF was achieved, as can be
seen in Figure 6(b). Some of the sand spilled over into the surrounding grass which will cause mixed materials when performing spectral unmixing. Therefore, when compiling the grass spectral reflectance, the grass region spectrally measured
was located approximately 2m away from the sand volleyball pit, thus reducing the amount of sand present in the spectral
measurements, but also keeping the type, color, and health of the grass even with the grass located right on the sand pit edge.
Proc. of SPIE Vol. 8743 87430G-4
Downloaded From: http://spiedigitallibrary.org/ on 08/07/2013 Terms of Use: http://spiedl.org/terms
(a)
(b)
(c)
Figure 4. Digital photos of all three targets during SHARE 2012 ground-truth collect. Figure 4(a) shows the black tarp deployed in the
volleyball pit, Figure 4(b) shows the white tarp deployed in the volleyball pit, and Figure 4(c) shows the gray tarp laid out over parking
lot gravel.
4. APPLICATIONS FOR SHARE LARGE TARGETS
There are many applications for which the large SHARE targets can be used. These include vicarious radiometric calibration, automatic gain correction, spectral unmixing, atmopheric adjacency effects and analysis, Modulation Transfer
Function calculation, and geo-correction to name a few. The following sections discuss how spectral unmixing validation
could use the available large target data as well as touching upon vicarious radiometric calibration. However, this is not the
limit to this dataset, as noted by the lengthy list of applications before.
4.1 Spectral Unmixing
Spectral unmixing methods identify and separate the constituent materials (or endmembers, EMs) within a single HSI
pixel. EM spectra and EM abundance maps are the results of spectral unmixing.8 Most spectral unmixing is based on a
linear mixture model where the pixel spectrum x is approximated as a linear combination of the N different EMs (e) and
their abundances (α),
N
X
x ≈ x̂ =
αi ei .
(1)
i=1
However, without a spectral library (or truth data) which associates the EM spectra to known materials, it is not possible
to convert the unmixed EM abundance maps into material abundance maps. Additionally, validating most of the steps
involved in spectral unmixing is a challenging topic, for which this SHARE dataset can be used.
Proc. of SPIE Vol. 8743 87430G-5
Downloaded From: http://spiedigitallibrary.org/ on 08/07/2013 Terms of Use: http://spiedl.org/terms
Background Materials Spectral Profiles
Grass
Sand
-,- Gravel
o
d 0.5
c>
C
0
400
900
1400
1900
2400
Wavelength [nmj
Figure 5. Average spectral profile of the three background materials for the three large targets during SHARE 2012.
During the identification and extraction of the EMs in spectral unmixing, the EMs are considered to be the pure
materials in the imagery. This is becuase the linear mixture model identifies the EMs as the corners of the convex hull and
all other pixels which fall within the hull corners are a combination of the EMs. However, how do we validate that the
extracted EM (or material distribution) is the correct pure pixel (or distribution). Using a small image chip extracted to
include the region around the black and white targets in the SHARE SpecTIR dataset (see Figure 5), it is easy to identify
the black tarp, white tarp, grass, and sand. Outlines of the pure material regions can be made by using the GPS data and
ground-truth photos, and for the black and white tarp it would seem logical to assume that the center of these tarps would
contain the purest pixel (or the endmember) for these two materials. Therefore, if endmember extraction routines are
applied to this image chip and do not pick out the center pixel of the black and white tarps further investigation needs to be
conducted on why these two approaches for determining single pixel purity do not agree (see Figure 7 where the SMACC
algorithm9 is used and three EMs are found not corresponding to the center of the tarps). Additionally, after finding the
EMs, statistical analysis can be applied to determine how similar the EMs are to the average ground-truth material spectra
or where the EM falls within the ground-truth spectral distribution of that material; in other words, how well does a single
pure EM represent the material distribution.
4.2 Vicarious Radiometric Calibration
Vicarious radiometeric calibration is a method for monitoring the operational status of a sensor. It also allows for multiple
sensors to be cross-calibrated to each other. However, to accomplish vicarious radiometric calibration, the absolute radiance
from a target must be known on the ground and the atmophere through which the radiance travels must also be wellbehaved, understood, and modeled. The black and white targets can be used to establish the brightest and darkest materials
in the Avon scene. As such the dynamic range of each imaging sensor can be stretched to fully span both of these targets.
Using the multiple temperal radiance measurements made of the large targets (one in the morning at 10:15am and a second
at 2:30pm), the full day’s worth of constant-panel full-sky radiance measurements (stationed near the large gray target), and
the two radiosonde data collections, the absolute radiance being reflected by the three large targets can be computed using
an atmospheric ray tracing program like MODTRAN.10 Additionally, these radiance values can ba propagated through the
atmosphere to compute top-of-the-atmosphere radiance values which can be compared to the different imaging sensors.
By this method the error in radiance that each sensor observed can be determined and cross-calibrated amounst each other.
Additionally, because of using multiple large targets (each with a different uniform reflectance value) any gain correction
needed to be applied to the sensors can be identified as well. The only draw back to the data provided is that the targets only
Proc. of SPIE Vol. 8743 87430G-6
Downloaded From: http://spiedigitallibrary.org/ on 08/07/2013 Terms of Use: http://spiedl.org/terms
(a)
(b)
(c)
Figure 6. Digital photos of the background materials for the large targets. Figure 6(a) shows the grass surrounding the volleyball pit,
Figure 6(b) shows the sand within the volleyball pit while it is being raked, and Figure 6(b) shows the gravel upon which the gray tarp
was placed.
cover a small region of the image and so the sensors’ non-uniformity can not be quantified. Similar vicarious calibration
approaches which have used large dry lake beds (instead of and in addition to large uniform tarps) have been applied to
several space-based and aerial-based sensors in the past, including Landsat 4-7, SPOT, AVIRIS, MODIS, MASTER, and
CHRIS.11–13
5. FUTURE WORK AND CONCLUSIONS
As discussed there are several tasks which can be performed with the large targets collected during SHARE 2012. At this
time the basic analysis of the data needed to extract useful image chips and perform quick-check calculations have been
completed. Orthorectification of the images needs to be performed before further analysis is conducted. A new spectral
unmixing validation approach also needs to be looked at with the ortho-rectified data which may make it possible to know
with better accuracy the true abundances of mixed pixels along the straight edges between the large targets. Additional
uniformity of the white and gray tarps needs to also be completed. Conversion of the ground-truth relative reflectance data
must also be converted to absolute reflectance by removing any spectralon panel spectral effects.
Ground-truth datasets like Indian Pines and Cuprite AVIRIS images are needed for algorithm development and validation. The SHARE 2012 dataset contains many different types of targets and materials for many different tests, algorithms,
experiments, etc. Here focus was on the large targets deployed in the volleyball pit and on the gravel parking lot within
the Avon, NY, park. Three tarps were deployed (black, white, and grey). The spectral and physical properties of each of
these tarps have been discussed. The black and white tarp were originally planned to be used for a novel spectral unmixing
validation algorithm, but have been determined to also be useful for other applications including vicarious radiometric
Proc. of SPIE Vol. 8743 87430G-7
Downloaded From: http://spiedigitallibrary.org/ on 08/07/2013 Terms of Use: http://spiedl.org/terms
(a)
(b)
(c)
77
56
34
'
(d)
Figure 7. Figure 7(a) shows RGB image of small HSI chip extracted from SpecTIR SHARE 2012 data, Figure 7(b) shows the four ROIs
user defined pure pixels, Figure 7(c) shows the three EMs extracted using SMACC not including the shade EM, and Figure 7(d) shows
the pixel distribution for the RGB HSI chip with 450nm (axis labeled as 34), 550nm (axis labeled as 56), and 750nm (axis labeled as
77).
Proc. of SPIE Vol. 8743 87430G-8
Downloaded From: http://spiedigitallibrary.org/ on 08/07/2013 Terms of Use: http://spiedl.org/terms
calibration, MTF calculation, atmospheric adjacency effects, and several others. As soon as possible all of this data will
be made available to the scientific community, at which point it becomes up to the many users what they can do with such
targets.
REFERENCES
1. Giannandrea, A., “The share 2012 data collection campaign,” submitted to Proc. SPIE 2013 .
2. Clark, R., Swayze, G., and Gallagher, A., “Mapping minerals with imaging spectroscopy.” Office of Mineral Resources Bulletin 2039 (November 1993). http://speclab.cr.usgs.gov.
3. Swayze, G., Clark, R., Sutley, S., and Gallagher, A., “Ground-truthing aviris mineral mapping at cuprite, nevada,” in
[Summaries of the third annual JPL airborne geosciences workshop. AVIRIS workshop. JPL Publication 192-14 ], 1,
47–49 (1992).
4. Zare, A. and Gader, P., “Hyperspectral band selection and endmember detection using sparsity promoting priors,”
IEEE Geoscience and Remote Sensing Letters 5, 256–260 (April 2008).
5. Kerekes, J. and Snyder, D., “Unresolved target detection blind test project overview,” Proceedings of SPIE 7695, 1–8,
Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XVI, Proceedings of SPIE
(2010).
6. Keshava, N., “Distance metrics and band selection in hyperspectral processing with applications to material identification and spectral libraries,” Geoscience and Remote Sensing, IEEE Transactions on 42(7), 1552–1565 (2004).
7. Goldberg, D. S., Kerekes, J. P., and Canham, K. A., “Hyperspectral linear unmixing: Quantitative evaluation of novel
target design and edge unmixing technique,” (2012).
8. Keshava, N. and Mustard, J. F., “Spectral unmixing,” Signal Processing Magazine, IEEE 19, 44–57 (January 2002).
9. Gruninger, J. H., Ratkowski, A. J., and Hoke, M. L., “The sequential maximum angle convex cone (SMACC) endmember model,” in [Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery X ],
Shen, S. S. and Lewis, P. E., eds., 5425(1), 1–14, SPIE (2004).
10. Schott, J., [Remote Sensing: an image chain approach ], Oxford University Press, 198 Madison Ave. New York, New
York 10016, second ed. (2007). ch. 9 and 10, pp. 371-479.
11. Naughton, D., Brunn, A., Czapla-Myers, J., Douglass, S., Thiele, M., Weichelt, H., and Oxfort, M., “Absolute
radiometric calibration of the rapideye multispectral imager using the reflectance-based vicarious calibration method,”
J. Appl. Remote Sens. 5(1) (2011).
12. McCorkel, J., Thome, K., Leisso, N., Anderson, N., and Czapla-Myers, J., “Radiometric characterization of hyperspectral imagers using multispectral sensors,” in [Proc. SPIE ], Gu, J. J. B. X. X. X., ed., 7452 (2009).
13. Czapla-Myers, J. S., Leisso, N. P., Anderson, N. J., and Biggar, S. F., “On-orbit radiometric calibration of earthobserving sensors using the radiometric calibration test site (radcats),” SPIE 8390 (April 2012).
Proc. of SPIE Vol. 8743 87430G-9
Downloaded From: http://spiedigitallibrary.org/ on 08/07/2013 Terms of Use: http://spiedl.org/terms