COMPARISON OF SPACEBORNE AND AIRBORNE HYPERSPECTRAL IMAGING

SYSTEMS FOR ENVIRONMENTAL MAPPING

Haluk Cetin

Mid-America Remote Sensing Center, Murray State University, Murray KY 42071 USA - [email protected]

KEY WORDS: Remote Sensing, Hyperspectral, Comparison, Platforms, Land Cover Mapping, Performance, Hyperion, AVIRIS

ABSTRACT:

The main purpose of this study was to compare hyperspectral remotely sensed data collected by the Hyperion satellite, and the
airborne Real-time Data Acquisition Camera System (RDACS-3) and the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS)
for environmental mapping and vegetation species identification. Hyperion was NASA's first hyperspectral imager aboard NASA's
Earth Observing-1 (EO-1) spacecraft. The EO -1 mission had three advanced land imaging instruments; Advanced Land Imager,
Hyperion, and Atmospheric Corrector. AVIRIS collects 224 contiguous spectral bands with wavelengths from 0.4 to 2.5 µm,
whereas RDACS-3 has many spectral modes (64, 128, 256, etc.). The study area, Land-Between-the-Lakes (LBL), is located in
western Kentucky, USA. Most of LBL consists of forested areas, which are predominantly oak and hickory, and open land areas.
Twenty-five percent (17200 hectares) of LBL falls within the Biosphere Reserve. AVIRIS was flown on the Twin Otter turboprop
at approximately 4000m above the ground with 4m spatial resolution on November 11, 1999 and September 10, 2001. The
Hyperion provided 242 spectral bands (from 0.4 to 2.5 µm) with a 30 meter spatial resolution and covered 7.5km by 200km area on
April 29, 2001. An RDACS-3 imagery with 120 spectral bands and 2x4m spatial resolution was collected at 2350m above the
ground by the ITD Spectral Visions on September 7, 1999. During the overflights, ground spectra using an ASD FieldSpec-FR®
spectroradiometer (0.35-2.5 µm) were collected for data calibration, spectral library construction, atmospheric correction and species
identification. Moreover, multispectral satellite and aerial imagery at 1m resolution was collected for some of the test sites in the
area. Several hyperspectral and multispectral processing tools were utilized for atmospheric corrections, enhancements, and
classifications. Best results were obtained using the AVIRIS and RDACS-3 data. The Hyperion data also provided very good
results for the mapping; however, its spatial resolution was one of the limitations of the Hyperion sensor. The statistical difference
among the classifications using the sensors proved to be mostly significant.

1. INTRODUCTION multispectral imagery could not be used for very detailed

mapping and identification of surface material, for which
The primary goal of the NASA Earth Observation System hyperspectral and/or ultraspectral sensors have been utilized.
(EOS) is to study the effects of climate on terrestrial Unlike the multispectral classifiers, hyperspectral classifiers
vegetation (Huete et al., 1994). The development of are used to identify objects using spectral endmembers in
multispectral imaging spectrometers during the early 1970's spectral libraries. Many attempts have been made to classify
allowed scientists for the first time to classify large areas of hyperspectral data using the traditional multispectral
terrain (Marmo, 1996). This led to the advent of classifiers. Classification time has been very long and
hyperspectral sensors with many bands and high spatial classification accuracy has not improved by the increased
resolution, allowing for the classification of large areas with number of bands when the multi-spectral classifiers were
finer spectral resolution (Cloutis, 1996). Current used (Lee and Landgrebe, 1993). Another approach using
multispectral satellites that orbit the earth have their own hyperspectral data has been mapping of cover types based on
limitations. The multispectral satellites such as Landsat and their abundances by using spectral unmixing techniques
SPOT as well as high spatial resolution sensors such as (Adams et al., 1986; Boardman 1990; Dwyer et al., 1995;
IKONOS and QuickBird have broad spectral bands. These Mustard and Pieters, 1987).
bands cover the visible, near and middle-infrared regions of
the electromagnetic spectrum (Jakubauskas and Price, 1997). Ecologists are now only beginning to explore the potential
This greatly reduces the ability of the multispectral sensor to uses of high spatial and high spectral resolution remote
spectrally discriminate between two objects on the ground sensing. For example, Schlesinger and Gramenopoulos (1996)
(Marmo, 1996). Multispecral sensors have been utilized for used high spatial resolution satellite imagery and aerial
many purposes including regional mapping. However, photography to test for desertification in the Sahel by
examining tree densities in images collected over 51 years. In draft Environmental Impact Statement for LBL identified
this study, no time-trend was observed, suggesting that if it is more than 20,000 acres to be designated as core areas in the
occurring at all, desertification is slower than previously preferred management alternative. The remainder of the
thought in the Sahel. A range of ecological problems become 170,000 acres of LBL serves as the buffer zone with the 17-
tractable with the possibility of locating and identifying county area is the transition zone. Many research studies
individual trees by remote sensing. At one end of the concerning LBL and Kentucky Lake have been conducted at
spectrum is the detection of rare individuals, genotypes, or Murray State University in the last 30 years; some of which
species and at the opposite end of the spectrum is the provided useful information for this research.
location and identification of individual trees of a common
species in a diverse community of similar species. A forester LBL is approximately 90 percent forested, which makes it
may have a particular interest in detecting the presence of rare one of the largest contiguous blocks of forested land east of
survivors of a disease or insect pest outbreak in order to find the Mississippi River in the US. Ferguson Spring/Energy
resistant individuals. For example, many field ecologists have Lake site has been selected as a research site because of its
observed occasional large American chestnut individuals that diverse ecology including wetland and bottom land/upland
have reached reproductive size and age despite exposure to forest. The dominant forest species are oak and hickory.
ubiquitous chestnut blight. It may be that 99 percent of such Oak and hickory trees requiring large amounts of sunlight
cases can be due to chance escape from the blight, but 1 when they are seedlings have been the dominant overstory
percent can be due to genetically-based resistance. Detection tree species in LBL. Forest composition of LBL would shift
of a large enough sample of reproductive chestnuts to toward maple and beech trees, which prefer shade if it is left
perform genetic screening could be impossible without an completely unmanaged. Maple trees do not produce nuts
extensive search procedure such as that provided by remote that wildlife can feed on maple seeds are small and winged.
sensing. This would negatively impact many different wildlife species,
including deer, squirrel, songbirds, and wild turkey, which for
The identification of species resisting to different stress thousands of years have fed on the acorns and nuts of
conditions has direct forestry and agriculture applications. oak/hickory forest of LBL. Table 1 summarizes some of the
The ability to identify vegetation at the species level using overstory species found in LBL.
hyperspectral data has been difficult because of lack of
information on vegetation characteristics and biochemical
characterization of vegetation at canopy level (Martin and
Aber, 1997).

There have been many studies comparing spaceborne and

airborne multispectral imagery. However, similar
comparisons could not be done for hyperspectral sensors
until the successful lunch of the Hyperion sensor.

The main objective of this study was to compare

hyperspectral remotely sensed data collected by the
Hyperion satellite, and the airborne Real-time Data
Acquisition Camera System (RDACS-3) and the Airborne
Visible/Infrared Imaging Spectrometer (AVIRIS) for
environmental mapping and vegetation species identification.

Figure 1. Location map of the study area and the datasets

2. STUDY AREA

The study area is located in the Land-Between-the Lakes The Center for Reservoir Research (CRR) was established in
(LBL) (Figure 1) National Recreation Area, which was 1987 by the Kentucky Council on Higher Education as a
designated in 1991 as an Internat ional Biosphere Reserve Center of Excellence for Teaching and Research. CRR
under the Man and Biosphere Programme of the United research has been focused mainly on Kentucky Lake water
Nations Education, Scientific and Cultural Organization monitoring and the Kentucky Lake GIS (KLGIS). One of the
(UNESCO). In addition to LBL, the Reserve includes 17 principal programs of CRR is the Kentucky Lake Long Term
surrounding counties in Kentucky and Tennessee as the area Monitoring Program (begun in July, 1988), which
influenced economically, socially and environmentally by the encompasses 17 primary monitoring sites on the lower 30 km
management of LBL. One of the main purposes of the of the lake. The KLGIS database providing a better
Biosphere Reserve Programme is to "involve industry, evaluation of the forest and wetlands in the area include
government, social agencies, schools, and special interest bathymetry, cultural features, geology, groundwater,
groups in management of the Reserve" and to encourage hydrography, soils, terrain, water quality and wetland data
cooperation in studying and solving regional problems. The for the Kentucky Lake drainage basin.
 Acronym Genus Species Common name with 120 spectral bands and 2x4m spatial resolution was
POHE Populus heterophylla swamp cottonwood collected at 2350m above the ground by the ITD Spectral
PRSE Prunus serotina black cherry Visions. Hyperion was NASA's first hyperspectral imager
aboard NASA's Earth Observing-1 (EO-1) spacecraft,which
QUAL Quercus alba white oak
had three land imaging instruments; Advanced Land Imager,
QUBI Quercus bicolor swamp white oak
Hyperion, and Atmospheric Corrector.
QUCO Quercus coccinea scarlet oak
QUFA Quercus falcata southern red oak 4. METHOD
QUIM Quercus imbricaria shingle oak
QULY Quercus lyrata overcup oak Hyperspectral imagery can be considered as a single image
dataset with a continuous spectrum of radiance (or
QUMA Quercus marilandica blackjack oak
reflectance) values associated with each image pixel (Bateson
QUMC Quercus macrocarpa bur Oak
and Curtiss, 1996). Hyperspectral imagery can distinguish
QUMI Quercus michauxii swamp chestnut between slope and brightness variations and resolve
oak absorption bands ni the spectrum, which can allow one to
QUMU Quercus muehlenbergii chinquapin oak identify surface material such as specific minerals or any
QUNI Quercus nigra water oak material with absorption features (Clark et al., 1992).
QUPA Quercus pagoda cherrybark oak AVIRIS was the first airborne hyperspectral sensor to
QUPH Quercus phellos willow oak measure reflected solar radiation from 400nm to 2500 nm
QUPL Quercus palustris pin oak (Green et al., 1998).
QUPR Quercus prinus chestnut oak
Individual bands of the RDACS hyperspectral datasets were
QURU Quercus rubra northern red oak
calibrated to percent reflectance using the known reflectances
QUSH Quercus shumardii shumard oak
of two gray scale placards placed on the ground during the
QUST Quercus stellata post oak overflight (Figure 4). The calibration and radiance to
QUVE Quercus velutina black oak reflectance conversions for the Hyperion dataset were done
RHCA Rhamnus caroliniana carolina buckthorn using several ground targets (dark, medium and light areas),
ROPS Robinia psuedoacacia black locust for which ground spectra were collected using the field
SAAL Sassafras albidum sassafras spectrometer. A simple linear regression model was used in
SANI Salix nigra black willow the calibration and conversion process.
TADI Taxodium distichum bald cypress
ULAL Ulmus alata winged elm
ULAM Ulmus americana american elm
ULRU Ulmus rubra slippery elm
Table 1. Summary of tree species found in LBL.

High spectral resolution (ultraspectral) leaf, soil, and water

reflectance data were collected on a regular basis with an ASD
Field Spec FR , a full-range field spectroradiometer, in the
field to create a spectral library to aid the classification of the
surface material. During the satellite and aerial data collection
events, ground spectra were also collected for selected
spectral targets.

3. IMAGERY

Several hyperspectral remotely sensed datasets collected by

the Hyperion satellite (Figure 2) on April 29, 2001, and the
airborne RDACS-3 (Figure 3) on on September 7, 1999 were
utilized in this research. AVIRIS data were also used but
because of some system problems with the sensor, only small
Figure 2. The Hyperion dataset , April 29, 2001
parts of the data were utilized.. AVIRIS was flown on the
Twin Otter turboprop at approximately 4000m above the
A spectral library of surface material (endmembers) and
ground with 4m spatial resolution on November 11, 1999 and
September 10, 2001. The Hyperion provided 242 spectral vegetation species (Table 1 and Figure 5) was created for
bands (from 0.4 to 2.5 µm) with a 30 meter spatial resolution hyperspectral analysis of the datasets. Several techniques
and covered 7.5km by 200km area. An RDACS-3 imagery
 Figure 5. Spectra of the vegetation species in LBL

provides a method for transforming multiple bands of data in

a predictable, and scene-independent way. These
Red: Band 110 (850nm),_Green: Band 60 (650nm), transformations may be designed so as to enhance a particular
Blue: Band 35 (550nm) cover-type, or to give the best visual representation of the
Figure 3. The RDACS dataset , September 7, 1999 multi-band image data. Spectral frequency plots of the nPDF
components give a spectral view of data distribution that can
including the Pixel Purity Index (PPI), an n-Dimensional be used to investigate the number and distribution of spectral
Visualizer, Spectral Angle Mapper (SAM) and Binary classes in a high dimensional data set. In addition, these plots
Encoding were utilized to map the study datasets. are used in a non-parametric classification of the image for
discrimination of discrete classes, as well as for classes that
For detailed classification, the n-Dimensional Probability are mixtures at the sub-pixel scale. In a mixed deciduous and
Functions (nPDF) approach was used (Cetin, 1990; Cetin coniferous forest an nPDF Deciduous Forest Index showed a
and Levandowski, 1991; Cetin et al., 1993). The nPDF high correlation with percent deciduous vegetation determined
techniques is an interactive from field surveys.

The nPDF approach may be explained using a cube model. A

generalized distribution of highly correlated digital remotely
sensed data in three dimensional feature space is shown in
Figure 1. In three-dimensional feature space the feature
vector is defined by X=[ x1,x2,x3]. The location of a point
within the range of the total possible measurement space can
be described by the distances to the two corners of the cube
shown in Figure 6. They are:
a b
2 2 2 1/2
D = (x + x + x ) , (1)
1 1 2 3
2 2 2 1/2
D = [ x + x + (R -x ) ] (2)
2 1 2 3

c For the multi-dimensional case, the feature vector is defined

e
d by X=[x1,x2,x3,...,xn], where n is the dimension of the data
Figure 4. a) Gray-scale placards, b) RDACS imafe showing and R is the maximum possible range of the data (255 for 8
the location of the placards, c) GPS measurement, d) bit data.) When a hyper-dimensional cube is used, the vector
sycamore leaves; healthy and under stress, and e) Full range magnitudes (the distances to the two corners) for n-
(350-2500nm) field spectroradiometer used in this study dimensional data are:

 n 2 1/2
image analysis technique, which overcomes many of the

D =  ∑ x 
inherent limitations of traditional classifiers. The techniques
has applications in three broad areas: data visualization, (3)
1 j
enhancement and classification. For data visualization, nPDF  j=1 
  n 2 1/2 nPDF i = S * Di / (2BIT * NB 1/2) (6)

D =  ∑ x * ( 1 - aj ) + (R - x ) * aj 
2
(4)
2 j j
 j=1 
where:

nPDF i = Component i of nPDF,

i = Corner number,
where j is the band number. A generalized formula for the
S = Desired scale for the nPDF axes,
distance to the corners of a hyper-dimensional cube can be
D = Calculated distance for component i,
written as (i is the corner or component number): i
BIT = Number of bits of input data,
NB = Number of bands used.

 n 2 1/2
D =  ∑ x * ( 1 - aj ) + (R - x ) * aj 
2 Frequency plots of two nPDF components (hyper -
(5)
i j j
 j=1 
dimensional distances) provide an excellent perspective of
multidimensional data distribution. Depending on the
spectral distribution of the classes of interest, the user can
select corners which provide the maximum separation of the
classes. A convenient scale for these nPDF components is 8
bit in range, and thus a two-dimensional frequency plot
requires a 256 by 256 array.

The cube model has the advantage of being a conceptually

simple way of describing corners in multidimensional space.
However, it does tend to limit the choice of corners for four
and higher dimensional data. Where this is a problem, the "a"
values (see equation 1) are used to describe the corner
location. Thus in Figure 6, corner #2 is also labeled (001),
which can be interpreted as a corner that has "a" value of zero
for the first two bands, and that of one in the third band.
Using this convention, the length of the list of "a" values
depends on the number of input bands, and thus the corner
corresponding to the origin in a four band image would be
described as (0000).

Prior to the classification process the spectral values for the

entire scene are transformed into nPDF space. The software
Figure 6. nPDF Cube Model
allows the user to view the distribution of the data and
There are eight possible corners of a three-dimensional cube enhance the data by interactively stretching and rotating.
as is shown in Figure 6. Four of the corners can be selected This allows a rough visual identification of separable classes.
as principal corners (1 through 4), the remaining corners (5 For the supervised classification procedure the training field
through 8) are the complimentary to the four principal data are then plotted into nPDF space. Polygons can then be
corners. For the hyper-dimensional cube model, "a" values drawn around the obvious classes to delineate the spectral
for the equation (5) are as follows (j is the band number): boundaries. The classification procedure uses the boundaries
of these polygons to assign pixels to the appropriate class.
D : For all j values a=0
1
5. RESULTS AND DISCUSSION
j = 1,2,4,5,… a = 0
D2 : if { j = 3,6,… a=1 Classification of the RDACS-3 dataset provided the highest
overall accuracy (76% overall accuracy for the overstory
j = 1,3,4,6,… a = 0 species classification and 94% overall accuracy for the other
D3 : if { j = 2,5,… a=1 land cover classes such as agriculture fields). The AVIRIS
j = 1,4,… a=0 datasets provided an overall accuracy of 69% for the
D4 : if { j = 2,3,5,6,… a=1 overstory species and 83% for the other land cover classes.
The Hyperion dataset provided 62% overall accuracy for the
The nPDF formula is: overstory species and 81% for the other classes (Figures 7
and 8; tones of red were used for the forest species).
A comparison between the classifications indicated that they
were significantly different. A Z statistic of 3.7 (Hyperion
and RDACS-3) and a Z statistic of 4.5 ((Hyperion and
AVIRIS) were computed for the pairwise comparisons.
Because the Z statistic values were greater than 1.96, there
were significant differences in the results of the classifications
of the sensors.

Figure 8. Classification map of the RDACS-3 dataset

Figure 9. Spectra of black willow; healthy and under stress

(the scale is between 0 and 1 (100%) for the reflectance axis)

6. CONCLUSIONS

Airborne and spaceborne hyperspectral imagery is becoming

increasingly accessible due to the increasing number of
companies and agencies operating hyperspectral scanners.
Airborne data acquisitions benefit greatly over satellite based
missions because the user has influence on the mission in
terms of time schedules, flight line arrangements, calibration
measurements, spectral/spatial resolutions, and acceptable
Figure 7. Classification map of the Hyperion dataset
weather conditions. However, airborne hyperspectral
sensors are often very expensive due to fact that limited
Another objective of this research was to outline forested
spatial coverage and multiple flight lines may be required to
areas that were under stress due to a drought. Figure 9 shows
cover a study area. Also, data processing is usually complex
the spectra of black willow leaves; healthy (cyan, green and
and can cause problems.
red lines), moderately stressed (blue) and severely stressed
(black). Another objective was to map the water quality in
Airborne hyperspectral sensors are usually used to test
the lakes using the hyperspectral data (Figure 10). However,
spaceborne hyperspectral sensors, which provide continuous
because of the page limitations, these studies were not
coverage of most of our planet as well planetary surfaces .
included in this paper.
 Cetin, H., Warner, T. A., and Levandowski, D. W., 1993.
Data classification, visualization and enhancement using n-
dimensional probability density functions (nPDF): AVIRIS,
TIMS, TM and geophysical applications. Photogrammetric
Engineering and Remote Sensing, 59(12): 1755-1764.

Clark, R.N., G.A. Swayze, and A. Gallagher, 1992. Mapping

t h e mineralogy and lithology of canyonlands, Utah with
imaging spectrometer data and the Multiple Spectral Feature
Mapping Algorithm. The Third Annual JPL Airborne
Geosciences Workshop, Volume I: AVIRIS Workshop; JPL
Publication, 92-14:11-13

Cloutis, E. A., 1996. Hyperspectral geological remote

Figure 10. High resolution spectra of Kentucky Lake (the sensing: evaluation of analytical techniques. International
scale is between 0 and 1 (100%) for the reflectance axis) Journal of Remote Sensing, 17(12):2215-2242.

The RDACS data used in this study provided the highest Dwyer, J. L., F. A. Kruse, and A. B. Lefkoff, 1995, Effects
accuracy in terms of classification of individual overstory of empirical versus model-based reflectance calibration on
species. Although AVIRIS generally provides very good automated analysis of imaging spectrometer data: A case
results, the datasets collected for the study area had many study from the Drum Mountains, Utah. Photogrammetric
problems. Only very limited areas could be used to classify Engineering and Remote Sensing, 61:1247-1254.
the land cover; therefore, the AVIRIS datasets were not fully
utilized in the comparison process. Although the Hyperion Green, R.O., M.L. Eastwood, C.M. Sarture, T.G. Chrien, M.
data had a low spatial resolution, the results showed that the Aronsson, B.J. Chippendale, J.A. Faust, B.E. Pauri, C.J.
data could be used for mapping of vegetation alliances in Chovit, M. Solis, M.R. Olah and O. Williams, 1998. Imaging
forestry related studies. Water quality studies using the Spectroscopy and the Airborne Visible/Infrared Imaging
Hyperion sensor should be more cost effective than using Spectrometer (AVIRIS). Remote Sensing of the Environment,
airborne hyperspectral imagery. 65:227-248.

Huete, A., Justice, C., and Liu, H., 1994. Development of

7. REFERENCES vegetation and oil indices for MODIS-EOS. Remote Sensing
of Environment, 49:224-234.
Adams, J. B., M. O. Smith, and P. E. Johnson, 1986.
Spectral mixture modeling: A new analysis of rock and soil
Jakubauskas, M.E., and Price, K.E., 1997. Empirical
types at Viking Lander 1 site. Journal of Geophysical
relationships between structural and spectral factors of
Research (JGR), 91:8098-8112.
Yellowstone Lodgepole pine forests. Photogrammetric
Engineering and Remote Sensing, 63(12):1375-1381.
Bateson, A., and B. Curtiss, 1996. A method for manual
endmember selection and spectral unmixing. Remote Sensing
Lee, C. and D. A. Landgrebe, 1993. Analyzing high -
of the Environment, 55:229-243.
dimensional multispectral data. IEEE Transactions on
Geoscience and Remote Sensing, 31 (4):792-800.
Boardman, J. W., 1990. Inversion of high spectral resolution
data. Society of Photo-Optical Instrumentation Engineers
M armo, J., 1996. Hyperspectral imager will view many
(SPIE), 1298:222-233.
colors of earth. Laser FocusWorld, 8:85-92.

Cetin, H., 1990. nPDF-An algorithm for mapping n-

Martin, M. E. and J. D. Aber, 1997. High spectral resolution
dimensional probability density functions for remotely
remote sensing of forest canopy lignin, nitrogen, and
sensed data. Proceedings of the 10th Annual International
ecosystem processes. Ecological Applications, 7 (2):431-443.
Geoscience & Remote Sensing Symposium, IGARSS'90, I:
353-356.
Mustard, J. F., and Pieters, 1987. Quantitative abundance
estimates from bidirection reflectance measurements. The
Cetin, H. and Levandowski, D. W., 1991. Interactive
17th Lunar Planet Science Conference, JGR, supplemen,:
classification and mapping of multi-dimensional remotely
92:E617-E626.
sensed data using n-dimensional probability density functions
(nPDF). Photogrammetric Engineering and Remote Sensing,
57(12): 1579-1587.
 8. ACKNOWLEDGMENTS

The author wish to thank the Scientific Data Purchase

Program, the Earth Science Applications Directorate at the
Stennis Space Center of NASA for providing the Hyperion
data and other satellite datasets of the study area, the ITD
Spectral Visions for providing the RDACS datasets and the
calibration placards used during the overflights, and JPL of
NASA for providing the AVIRIS datasets. This research was
funded by a Kentucky NASA -EPSCoR Research Grant
(NCC5-571) and an NSF-EPSCoR Research Grant (EPS-
0132295).