You are on page 1of 10

Comparison between empirical and physically based models

of atmospheric correction
E. Mandanici a*, F. Franci a, G. Bitelli a, A. Agapiou b, D. Alexakis b and D. G. Hadjimitsis b
a
Dept. of Civil, Chemical, Environmental and Materials Engineering (DICAM) –
University of Bologna, viale Risorgimento 2, Bologna, Italy; b Department of Civil Engineering and
Geomatics Cyprus University of Technology, Address, Cyprus.

ABSTRACT

A number of methods have been proposed for the atmospheric correction of the multispectral satellite images, based on
either atmosphere modelling or images themselves. Full radiative transfer models require a lot of ancillary information
about the atmospheric conditions at the acquisition time. Whereas, image based methods cannot account for all the
involved phenomena.
Therefore, the aim of this paper is the comparison of different atmospheric correction methods for multispectral satellite
images. The experimentation was carried out on a study area located in the catchment area of Yialias river, 20 km South
of Nicosia, the Cyprus capital. The following models were tested, both empirical and physically based: Dark object
subtraction, QUAC, Empirical line, 6SV, and FLAASH. They were applied on a Landsat 8 multispectral image.
The spectral signatures of ten different land cover types were measured during a field campaign in 2013 and 15 samples
were collected for laboratory measurements in a second campaign in 2014. GER 1500 spectroradiometer was used; this
instrument can record electromagnetic radiation from 350 up to 1050 nm, includes 512 different channels and each
channel covers about 1.5 nm. The spectral signatures measured were used to simulate the reflectance values for the
multispectral sensor bands by applying relative spectral response filters. These data were considered as ground truth to
assess the accuracy of the different image correction models.
Results do not allow to establish which method is the most accurate. The physics-based methods describe better the
shape of the signatures, whereas the image-based models perform better regarding the overall albedo.
Keywords: atmospheric correction models, Yialias river (Cyprus), spectroradiometer measurements

1. INTRODUCTION
Atmospheric correction is one of the most troublesome problem in remote sensing, since the distributions and intensities
of atmospheric effects are often inadequately known. A number of methods have been proposed for the correction of
multispectral satellite images, based on either atmosphere modelling or images themselves. Despite the variety of
techniques used to estimate the atmospheric effect, the atmospheric correction remains a hard task in the pre-processing
of image data2,7.
The aim of this paper is the comparison of different atmospheric correction methods for multispectral satellite images.
The experimentation was carried out on a study area located in the portion of the Yialias river basin between
Lythrodonta and Dhali villages, in the central part of the Cyprus island. Specifically, this area is situated in the broader
region of Nicosia, between longitudes 33° 16′ 8″ and 33° 24′ 1″ and latitudes 34° 56′ 54″ and 35° 1′ 41″. It is mostly
characterized by the presence of agricultural areas (olive and carob trees, fruit trees, berry plantations and cereals), scrub
vegetation and coniferous forests. From West to East metamorphic and volcanic rocks outcrop followed by sedimentary
units such as marl-chalk3.
In order to evaluate the performance of different correction models, the spectral signatures of ten land cover types were
measured during a field survey carried on in 2013. This campaign was performed in the framework of the
“SATFLOOD” project13, implemented by the Department of Civil Engineering and Geomatics of the Cyprus University
of Technology (CUT) for the integrated use of satellite remote sensing and hydraulic modeling for flood risk assessment

*
emanuele.mandanici@unibo.it ; phone 39 051 2093107; fax 39 051 2093114; dicam.unibo.it.

Third International Conference on Remote Sensing and Geoinformation of the Environment (RSCy2015),
edited by Diofantos G. Hadjimitsis, Kyriacos Themistocleous, Silas Michaelides, Giorgos Papadavid,
Proc. of SPIE Vol. 9535, 95350E · © 2015 SPIE · CCC code: 0277-786X/15/$18 · doi: 10.1117/12.2193176

Proc. of SPIE Vol. 9535 95350E-1

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 01/14/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx


at catchment scale in Cyprus3. Moreover, 15 samples for five different soil types were collected for laboratory
measurements during a second campaign in 2014.

33°160E 33°180E 33°200E 33°22'0E 33°240E


35'2'0"N 35°2'0"N

35'0'0"N 35°0'0"N

34580N 34°58'0"N

33160E 33°180E 33°20'0"E 33220E 33°24'0E

ground measurements 2013 A samples collection 2014

Figure 1 - Location of the study area. The markers highlight the ground measurements and soil sample collection positions.

2. ATMOSPHERIC CORRECTION METHODS


Following the scheme proposed by Hadjimitsis and Clayton7, the atmospheric correction methods can be divided in
absolute vs. relative corrections, the former lead to surface reflectance values, the latter aim only at normalizing
multitemporal images. The present paper focuses on absolute corrections only. They can be categorized in image-based
corrections, which use data derived from statistical analyses of images themselves, and physics-based corrections, which
exploit an implementation of a radiative transfer model but require external data. In the following paragraphs the
distinctive aspects of the algorithms tested in the present work are summarized.
2.1 Dark Object Subtraction (DOS)
This is, maybe, one of the oldest and simplest methods proposed in literature8. The basic idea is finding a surface whose
reflectance is so low that its contribution to the signal recorded by a sensor is negligible compared with the radiance
diffused by the atmosphere. The minimum digital number of the pixels corresponding to this surface can be subtracted to
all the pixels of the image12. To obtain reflectance values, the image can be calibrated to top of atmosphere (TOA)
reflectance, before applying the DOS. Since it consists in a mere subtraction, this method can mitigate only the
“additive” contribution due to the scattering, but it cannot compensate for the atmospheric absorption.
2.2 Empirical Line (EL)
The empirical line method relies on the collection of the reflectance spectra of a few targets on the ground. These targets
should be chosen in areas as spectrally homogeneous as possible for an extension of at least three times the image pixel
size. This criterion is not easy to meet, especially for medium resolution imagery. Furthermore, the targets should be
selected carefully to cover the whole range of reflectances9. Given the surface reflectance (ρ) of the targets, a regression

Proc. of SPIE Vol. 9535 95350E-2

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 01/14/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx


is computed to establish a linear relation with the image radiance (L) of the corresponding pixels in the image, in the
following form:
L− p
ρ= (1)
I
where the additive term p can be interpreted as the path radiance and the multiplicative term I as the contribution of the
solar irradiance and the absorption. Clearly, the surface reflectance curve is to be resampled to simulate the spectral
bands of the imaging sensor.
2.3 QUick Atmospheric Correction (QUAC)
QUAC is a semi-empirical atmospheric correction method, which also enables retrieval of the wavelength-dependent
optical depth of the aerosol or haze and molecular absorbers4. Differently from physics-based algorithms, it determines
this atmospheric compensation parameters directly from the information contained within the scene, without any
ancillary information. The retrieval is based on the following linear equation:

ρTOA = A + Bρ + Cρ e (2)

where ρTOA is the apparent reflectance (top of atmosphere), ρe is a spatially averaged surface reflectance and A, B and C
are coefficients that describe the transmission and scattering effects of the atmosphere. To provide an accurate solution,
the model requires that at least ten different materials and a sufficient number of dark pixels be present within the
image4.
2.4 Second Simulation of a Satellite Signal in the Solar Spectrum Vector code (6SV)
This model is a basic radiation transfer code, originally developed for the calculation of look-up tables in the MODIS
(MODerate resolution Imaging Spectroradiometer) atmospheric correction algorithm10. This code is based on the method
of successive orders of scattering approximations and it includes the calculation of four components of the Stokes vector,
in order to compute the effect of polarization. Furthermore the last release provides the possibility to vary arbitrarily the
vertical aerosol profile and a more accurate calculation of aerosol scattering phase functions. The basic equation of the
model is very similar to the equation 3 proposed in the following section. For the present study an implementation of
code, called Visual SixS, was used to perform the atmospheric correction of the images11. It provides a graphical user
interface for the code and expands the sensor support, including OLI, WorldView-2 and Hyperion.
2.5 Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH)
This model was developed by the Air Force Research Laboratory, Space Vehicles Directorate, Hanscorn AFB and
Spectral Sciences, Inc., and it is now distributed as a package for ENVI6. It provides an accurate physics-based
derivation of atmospheric properties based on Modtran calculations and adds an image-sharpening ‘adjacency’
correction, with an internal ‘polishing’ algorithm. It provides also algorithms for column water vapor and aerosol
retrievals from data. The computation of surface reflectance ρ is based on the following equation:
Aρ Bρ e
L= + + Lp (3)
1 − Sρ e 1 − Sρ e
where ρe is an average surface reflectance for the surrounding region, S is the spherical albedo of the atmosphere, Lp is
the radiance backscattered by the atmosphere, and A and B are surface independent coefficients that vary with
atmospheric conditions and are computed from Modtran outputs6.

3. MATERIALS AND METHODS


The analyses performed here make use of ten spectra of different land cover types (Table 1) measured during the field
campaign performed on June 19, 2013 (within the framework of “SATFLOOD” project), plus five soil spectra measured
in the laboratory on 15 soil samples (marl-chalk, basalt and soils from cultivated lands), collected during the second
campaign on July 1, 2014 (Table 2).

Proc. of SPIE Vol. 9535 95350E-3

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 01/14/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx


The positions of in situ measurements and sample collection were recorded with Global Positioning System (GPS). For
both field campaign and laboratory measurements, GER 1500 spectroradiometer was used. This instrument can record
electromagnetic radiation from 350 up to 1050 nm; it includes 512 different channels and each channel covers about 1.5
nm. The field of view (FOV) of the instrument was set to 4°. The accuracy of radiometric measurements provided by the
manufacturer are: 400 nm ± 5%, 700 nm ± 4%, and 1000 nm ± 5%.

Table 1 - Locations of the field spectral measurements of 2013 campaign and description of the materials.
ID Description Longitude Latitude
F1 Ploughed field 33°24'05,882" 34°59'21,980"
F2 Chopped barley 33°24'05,911" 34°59'21,998"
F3 Marl-chalk 33°23'57,034" 35°00'19,739"
F4 Court 33°23'34,058" 35°01'12,018"
F5 Olives 33°23'57,037" 35°01'21,821"
F6 Cultivation (Alfa-Alfa) 33°24'02,804" 35°01'29,204"
F7 Low bushes 33°23'45,528" 35°00'25,618"
F8 Ploughed field 33°21'06,631" 34°58'59,002"
F9 Asphalt 33°20'05,896" 34°57'41,242"
F10 Volcanic rock 33°16'10,567" 34°56'55,435"

Table 2 - List of the soil samples collected for laboratory measurements during the 2014 campaign. Three samples were
collected in each site.
ID Description Longitude Latitude
S1-1 33° 23’ 11” 35° 00’ 07”
S1-2 Ploughed field 33° 23’ 11” 35° 00’ 06”
S1-3 33° 23’ 11” 35° 00’ 06”
S2-1 33° 21’ 43” 34° 59’ 20”
S2-2 Ploughed field 33° 21’ 44” 34° 59’ 20”
S2-3 33° 21’ 44” 34° 59’ 20”
S3-1 33° 21’ 06” 34° 58’ 59”
S3-2 Ploughed field 33° 21’ 07” 34° 58’ 59”
S3-3 33° 21’ 07” 34° 58’ 59”
S4-1 33° 19’ 53” 34° 58’ 11”
S4-2 Volcanic rock 33° 19’ 53” 34° 58’ 11”
S4-3 33° 19’ 54” 34° 58’ 12”
S5-1 33° 23’ 55” 35° 00’ 18”
S5-2 Marl-chalk 33° 23’ 56” 35° 00’ 18”
S5-3 33° 23’ 56” 35° 00’ 18”

Proc. of SPIE Vol. 9535 95350E-4

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 01/14/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx


Figure 2 - Some operation during 2014 field campaigns and laboratory measurements.
All the analyses were performed on a Landsat 8 OLI multispectral image acquired on July 5, 2014. The processing
involved only a subset of the entire image. The image was firstly calibrated to radiance, applying the gains and offsets
provided with metadata. Then, all the algorithms described in Section 2 were applied.
Radiative transfer models require ancillary data to describe properly the atmospheric conditions at the time of the image
acquisition. Since detailed vertical profiles of atmosphere were not available for the area, data derived from the
“Giovanni” online data system were used. This portal is developed and maintained by the NASA Goddard Earth
Sciences Data and Information Services Center1. In particular, total precipitable water vapor, total ozone burden and
aerosol optical thickness at 550 nm were derived from MODIS data available on Giovanni for the acquisition date (Table
3).

Table 3 - Atmosphere parameters obtained from NASA’s portal “Giovanni”.


Parameter Value MODIS product
Total precipitable water vapor (cm) 2.204 MOD08 Daily
Total Ozone burden (Du) 328.0 MOD08 Monthly
Optical thickness (at 550 nm) 0.278 MOD08 Daily

For the empirical line method, the spectra of two samples were used, i.e. S1 and S3, which are expected to be
representative of two large homogeneous areas. All the other spectra were used to compare the performance of the
atmospheric correction methods. The spectra acquired in the field were convolved with the relative spectral response
function of the OLI sensor. In particular, the first five bands of the Landsat image (Coastal, Blue, Green, Red, NIR) were
simulated, because they are the only bands within the spectral range of the GER 1500 spectroradiometer. Since three
samples were collected for each soil type in the 2014 campaign, the average spectral signatures were calculated and
considered for the analyses.
In order to perform comparisons between the reference field spectra and the signatures of the corresponding image
pixels, some indicators were computed: the Euclidean distance (ED), the spectral angle mapper (SAM) and the slope of
the regression line (SR). ED and SAM are defined in a five-dimensional “feature space”, whose dimensions are the five
OLI bands considered. RS, instead, is computed as the slope coefficient of the regression line approximating five points
(one per band), whose abscissae are defined by the reflectance value predicted by a model and ordinate by the values
measured by the spectroradiometer.

4. RESULTS AND DISCUSSION


The signatures of the pixels corresponding to the positions of the field measurements were extracted from the Landsat 8
image. This image was corrected with all the five methods mentioned above, thus, five signatures for each location can

Proc. of SPIE Vol. 9535 95350E-5

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 01/14/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx


be compared,, one for each model. Figurre 3 shows thee comparison of o the field sppectrum and thhe signatures extracted
e from
m
the image foor the F8 sam mple (Table 1)), acquired onn a ploughed d field. For thhe same sampple, the constrruction of thee
regression linnes for the com
mputing of thee SR indicatorr is shown in Figure
F 4.
Remarkable differences
d caan be observedd among the models
m appliedd. The two phhysics-based m models (FLAA ASH and 6SV)
are in good agreement
a betw ween each othher, but providde very different results commpared to thee image-based d models. Bothh
the physics-bbased methods and image-bbased ones reesult in signifficantly differrent values frrom the refereence spectrum m
derived fromm ground meaasures. In thhe example shhown in Figure 3, the physics-based methods oveerestimate thee
reflectance inn all the specctral bands, whereas
w the empirical
e methhods underesttimate the refflectance, esp pecially in thee
visible range.. However, coonsidering thee entire collecction of samplles, it is not possible to deffine a general trend. In factt,
for some sammples (e.g. F7) all the models overestimate the reflecttance, and forr other samplees (e.g. F4, S5) all of them m
underestimatee the reflectannce.

Figure 3 - Comparison of spectral signaatures obtained using different atmospheric coorrection modells with the conv volved
specttrum measured in the field (blaack line with veertical error barrs). Signatures refer
r to a ploughhed field at F8 site.

Proc. of SPIE Vol. 9535 95350E-6

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 01/14/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx


Figure 4 - Regression linne between the reflectance
r valuues of the field spectra and of the
t image pixell in all the band
ds for the
samee site of Figure 3. Black line reepresents the ideal slope (SR=1) and is reportted as referencee.
To analyze differences
d between spectraal signatures, two separate aspects can be b distinguishhed: the magn nitude and thee
t overall albedo of the surface,
shape. The fiirst refers to the s whilst the second to the relativee differences in reflectancee
among bandss. The indicaators computeed try to highhlight both th hese aspects. The ED is strongly depeendent on thee
difference in magnitude; thhe SAM is less sensitive too it and it is considered
c inssensitive to illlumination conditions12; thee
SR is theorettically indepenndent of the magnitude,
m if the
t gap betweeen the profilees were consttant in all the bands, the SR R
value would beb one. Thus, SAM and SR R are good indicators of shap pe similarity.
The ED, SAM M and SR inddicators for alll the samples measured in laboratory
l aree reported in F
Figure 5. The physics-basedd
methods (6SV, FLAASH H) exhibit sim t indicatorss (slightly better results in
milar performaance in all the n average forr
FLAASH). Both
B perform better
b than thee image-basedd methods if SAM
S and SR are considereed; conversely y they perform
m
worse considdering the EDD. This probabbly means thaat physics-bassed models arre more accurrate in term ofo shape, evenn
showing a rellevant gap.

Proc. of SPIE Vol. 9535 95350E-7

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 01/14/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx


Figure 5 - ED, SAM andd SR indicators plotted for all the
t samples of thet 2014 campaaign measured iin lab and for alll the
correection models. Black
B bars referr to the mean of all the samplees for each moddel.
Image-based methods (DO OS, EL) show a similar behaavior in the ED D indicator, whereas
w they ddiffer in SAM and SR. DOS S
performs bettter in the SR than in SAM M. This is probbably due to the t fact that DOS
D is a merre subtractionn of a constannt
value and thee SR indicator is insensitivee to additive coonstants, whille SAM is inseensitive to connstant multipllicative factorss
for each bandd. The resultss of the QUA AC model aree very poor co onsidering SA
AM and SR, meaning a co omplete shapee
mismatch. Thhis is evident especially
e in visible
v a by Bitelli and Mandaniici5.
bands, as observed also
Despite all thhese considerrations, the reesults obtainedd in the preseent work do not allow thee identification of the mosst
accurate moddel. There are several factors that can coompromise thee comparison between the reference speectra measuredd
on the field annd the signatuures extracted from the atmoospheric correected images:
• Pixeel size: each pixel
p of a Lanndsat image covers an areaa of 900 m2, so different m materials are likely to occurr;
convversely, a fieldd measure is representativee of a single material.
m This problem can be mitigated executing thee
measurements in sites where a uniform
u land cover type strretches for a laarge area.

Proc. of SPIE Vol. 9535 95350E-8

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 01/14/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx


• Resampling: Landsat images are distributed in the L1T processing level, meaning that they are geometrically
corrected using a digital elevation model and a collection of ground control points from a global dataset.
Unfortunately, the resampling is made by a cubic convolution, which involves a 4x4 kernel and thus alters
significantly the radiometry of the pixels, acting as a low-pass filter.
• Georeferencing: the medium spatial resolution of Landsat image (30 m) does not allow a precise coregistration
with the ground data. The eventual shift can introduce an error if there is a change in cover type. Also in this
case the choice of the site is important.
• Atmosphere conditions: for the physics-based methods, the input parameters describing the atmospheric
conditions are critical for the final accuracy; unfortunately, they are often affected by large uncertainty.
• Non-linearity: all the considered models, both physics-based and image-based, apply the correction to the
radiance image using a linear relation. This relation cannot account for the nonlinear components of the
atmospheric effects.
Further investigations may clarify the entity of the errors introduced by each source.

5. CONCLUSIONS
The paper presents a comparison of five atmospheric correction models, 6SV, DOS, Empirical line, FLAASH, QUAC,
applied on a Landsat 8 multispectral image, acquired on Cyprus island. The performances of these models were
evaluated using as reference 15 spectra measured in the field or in laboratory by a GER 1500 spectroradiometer. In order
to measure the similarity among the spectral signatures, three indicators were computed: the Euclidean distance, the
spectral angle mapper and the slope of the regression line (between the reflectance values of the field spectra and of the
image pixel in all the bands).
The obtained results show a high variability and do not allow to establish which method is the most accurate. The
physics-based methods describe better the shape of the signatures, whereas the image-based models (DOS, EL) perform
better regarding the overall albedo.
Further researches are required for a better understanding of the phenomena which can introduce spurious errors in the
comparison, such as material mixing due to the pixel size, resampling algorithms, or eventual shifts due to a non-perfect
coregistration with the ground data.

REFERENCES

[1] Acker, J., G., Leptoukh, G., “Online analysis enhances use of NASA Earth science data,” Eos Trans. AGU
88(2), 14-17 (2007).
[2] Agapiou, A., Hadjimitsis, D. G., Papoutsa C., Alexakis, D. D., Papadavid, G., “The Importance of Accounting
for Atmospheric Effects in the Application of NDVI and Interpretation of Satellite Imagery Supporting
Archaeological Research: The Case Studies of Palaepaphos and Nea Paphos Sites in Cyprus,” Remote Sensing
3, 2605-2629 (2011).
[3] Alexakis, D., D., Hadjimitsis, D., G., Agapiou, A., “Integrated use of remote sensing, GIS and precipitation data
for the assessment of soil erosion rate in the catchment area of ‘Yialias’ in Cyprus,” Atmospheric Research 131,
108–124 (2013).
[4] Bernstein, L., S., Adler-Golden, S., M. and Sundberg, R., L., “Validation of the QUick Atmospheric
Correction(QUAC) algorithm for VNIR-SWIR multi- and hyperspectral imagery,” Proc. SPIE 5806, 668-678
(2005).
[5] Bitelli G. and Mandanici E., “Atmospheric correction issues for water quality assessment from Remote Sensing:
the case of Lake Qarun (Egypt),” Proc. SPIE 7831, 78311Z (2010).
[6] Cooleya, T., Andersona, G., P., Felde, G., W., Hoke, M., L., Ratkowskia, A., J., Chetwynd, J., H., Gardner, J.,
A., Adler-Golden, S., M., Matthew, M., W., Berk, A., Bernstein, L., S., Acharya, P., K., Milled, D., Lewise, P.,
“FLAASH, a MODTRAN4-based Atmospheric Correction Algorithm, Its Application and Validation,”
Geoscience and Remote Sensing Symposium, 2002. IGARSS '02. IEEE International 3, 1414-1418 (2002).

Proc. of SPIE Vol. 9535 95350E-9

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 01/14/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx


[7] Hadjimitsis, D., G., and Clayton, C., R., I., “An assessment of the effectiveness of atmospheric correction
algorithms trough the remote sensing of some reservoirs,” Int. J. Remote Sensing 25(18), 3651-3674 (2004).
[8] Hadjimitsis, D., G., Clayton, C., and Retails, A., “On the darkest pixel atmospheric correction algorithm: A
revised procedure applied over satellite remotely sensed images for environmental applications,” Proc. SPIE
5239, 464-471 (2004).
[9] Karpouzli, E., Malthus, T., “The empirical line method for the atmospheric correction of IKONOS imagery,”
Int. J. Remote Sensing 24(5), 1143-1150 (2003).
[10] Kotchenova, S., V., Vermote, E., F., Matarrese, R. and Klemm, F., J., “Validation of a vector version of the 6S
radiative transfer code for atmospheric correction of satellite data. Part I: Path radiance,” Applied Optics 45(26),
6762-6774 (2006).
[11] Mandanici, E., “Implementation of Hyperion sensor routine in 6SV radiative transfer code,” Proc.
Hyperspectral Workshop ESA SP-683, (2010).
[12] Mather, P., M., [Computer Processing of Remotely-Sensed Images], Third Edition, John Wiley & Sons (2004).
[13] SATFLOOD, “Project website”, http://blogs.cut.ac.cy/satflood/ (9 March 2015).

Proc. of SPIE Vol. 9535 95350E-10

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 01/14/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

You might also like