3:/6432+38'1"89*/+7')918< 9(1/)'8/437 3:/6432+38'1"89*/+7
A Comparison of AVIRIS and Landsat for Land
Use Classi#cation at the Urban Fringe
Rutherford V. Pla$
Geysburg College
Alexander F.H. Goetz
4114;8./7'3*'**/8/43'1;4607'8 .B57)9541'-+B<7(96-+*9+7,')
'684,8.+ 3:/6432+38'13*/)'8467'3*25')877+772+38422437
Share feedback about the accessibility of this item.
?/7/78.+59(1/7.+67:+67/434,8.+;460?/759(1/)'8/43'55+'67/3+B<7(96-411+-+7/378/898/43'16+547/846<(<5+62/77/434,
8.+)45<6/-.84;3+6,465+6743'197+348,466+*/786/(98/439541'5+62'3+381/30.B57)9541'-+B<7(96-+*9+7,')
?/745+3'))+77'68/)1+/7(649-.884<49(<?+9541'").41'67./5'8+B<7(96-411+-+8.'7(++3'))+58+*,46/3)197/43(<'3
'98.46/=+*'*2/3/786'8464,?+9541'46246+/3,462'8/4351+'7+)438')8 )9541'-+B<7(96-+*9
1'B!%4+8=425'6/7434,%!"'3*'3*7'8,46'3*$7+1'77/@)'8/43'88.+$6('36/3-+
.484-6'22+86/)3-/3++6/3-!+248+"+37/3-
A Comparison of AVIRIS and Landsat for Land Use Classi#cation at the
Urban Fringe
Abstract
38./7789*<;+8+78+*;.+8.+6%!"*'8''114;+*,46/2564:+*1'3*97+)1'77/@)'8/434:+67<38.+8/)
'3*7'8#*'8',46'14)'8/43438.+96('3696'1,6/3-+4,4146'*4A+6564)+77/3-8.+%!"/2'-+
'3*)6+'8/3-'7<38.+8/)'3*7'8/2'-+;+97+*78'3*'6*)1'77/@)'8/43'3*5478)1'77/@)'8/43564)+*96+784
)425'6+8.+*'8'7496)+7,461'3*97+2'55/3-&+,493*8.'8,468./714)'8/43%!".41*724*+78(98
6+'1'*:'38'-+74:+6'3*7'8,468.+)1'77/@)'8/434,.+8+64-+3+497'3*:+-+8'8+*1'3*97+7968.+6246+
8./7'*:'38'-+)42+7'12478+38/6+1<,6428.+1'6-+392(+64,7+374675+)86'1('3*76'8.+68.'38.+./-.
"/-3'1844/7+!'8/4"!
Keywords
%!"'3*78'8#4146'*4$6('36/3-+"/-3'1844/7+!'8/4
Disciplines
3:/6432+38'13*/)'8467'3*25')877+772+38>3:/6432+38'1")/+3)+7
?/7'68/)1+/7':'/1'(1+'8?+9541'").41'67./5'8+B<7(96-411+-+ .B57)9541'-+B<7(96-+*9+7,')
Abstract
In this study we tested whether AVIRIS data allowed for im-
proved land use classification over synthetic Landsat
ETM+
data for a location on the urban-rural fringe of Colorado. After
processing the
AVIRIS image and creating a synthetic Landsat
image, we used standard classification and post-classification
procedures to compare the data sources for land use mapping.
We found that, for this location,
AVIRIS holds modest, but real,
advantages over Landsat for the classification of heteroge-
neous and vegetated land uses. Furthermore, this advantage
comes almost entirely from the large number of sensor spec-
tral bands rather than the high Signal-to-Noise Ratio (
SNR).
Introduction
In rapidly urbanizing areas, such as the Front Range of Col-
orado, maps fast lose their validity. Large areas of prairie or
farmland land can be overrun by residential development in a
matter of months. Remotely sensed data allows land use and
land cover to be mapped quickly, relatively cheaply, and fre-
quently. With improved mapping of rapidly changing areas,
planners will be able to better address issues associated with
urban sprawl. However, the choice of sensor can significantly
influence the accuracy of the classification. While it is com-
monly thought that smaller Ground Sampling Distance (
GSD),
also called pixel size, is the key to better land use classifica-
tion, the number of spectral bands and the Signal-to-Noise
Ratio (
SNR) may influence classification accuracy as well.
Commonly, researchers use sensors such as those on
Landsat or
SPOT satellites for mapping land use and land cover
(Table 1). Of these, the Landsat sensors have more spectral
bands and a longer time series, while
SPOT provides smaller
GSD. Less traditional sensors may provide additional informa-
tion that can improve mapping accuracy. The Airborne Visible
Infrared Imaging Spectrometer (
AVIRIS), for example, produces
images with 224 spectral bands between 0.4 and 2.45 m,
compared to six bands for Landsat (not including the thermal
band) and three for
SPOT’s Multispectral Imager (XS). Sensors
with a large number of continuous spectral bands, such as
AVIRIS, are called hyperspectral imagers (Green et al., 1998).
Though hyperspectral imagers have been used in studies
of mineralogical mapping and ecology, they have rarely been
employed for land use mapping. A small number of studies
have explored the integration of hyperspectral and Synthetic
Aperture Radar (
SAR) for urban mapping (Gamba and Housh-
A Comparison of AVIRIS and Landsat for
Land Use Classification at the Urban Fringe
Rutherford V. Platt and Alexander F.H. Goetz
mand, 2001; Hepner et al., 1998). Other studies have used
hyperspectral imagery to map a narrow range of urban materi-
als and processes (Ben-Dor et al., 2001; Ridd et al., 1992; Salu,
1995). One study used an iterative spectral un-mixing proce-
dure to delineate urban materials (Roessner et al., 2001). To
date, however, no studies have tested whether hyperspectral
imagery improves land use classification accuracy over and
above multispectral imagery such as from Landsat.
In this study, we tested whether
AVIRIS data allowed for
improved land use classification over synthetic Landsat ETM+
data for a location on the urban-rural fringe of Colorado. We
expected that the large number of bands and high
SNR pro-
vided by AVIRIS would help distinguish land cover types that
are easily confused (irrigated urban areas and irrigated crops,
for example). After processing the
AVIRIS image and creating a
synthetic Landsat image, we used standard classification and
post-classification procedures to compare the data sources for
land use mapping.
Sensor Specifications and Classification Accuracy: The Case
of the Urban Fringe
Among the factors that may influence classification accuracy
are the Ground Sampling Distance (
GSD), number of spectral
bands, and Signal-to-Noise Ratio (
SNR) of a sensor. Generally, it
is thought that
GSD is the most important factor for classification
accuracy of built environments (Forster, 1985). For example, a
study in Indonesia found that
SPOT Multispectral (XS) images are
superior to Landsat Multispectral Scanner (
MSS) images for
mapping of heterogeneous, near-urban land cover because of
SPOT’s smaller pixel size (Gastellu-Etchegorry, 1990). The link
between
GSD and classification accuracy, however, is sometimes
tenuous. In heterogeneous areas, such as residential areas, it has
been shown that classification accuracies may actually improve
by up to 20 percent as
GSD is increased (Cushnie, 1987). This oc-
curs when the reflectance spectra of a variety of cover types in
an urban environment blend to form an overall urban signal that
can be easily distinguished from other land covers.
SNR, which varies sensor-by-sensor and band-by-band and
pixel-by-pixel, may also influence classification accuracy. The
greater the
SNR, the more usable information is available in
the data. Overall,
AVIRIS has much higher SNR than Landsat
sensors. Within the Landsat family, the
ETM+ in Landsat 7 has
a higher
SNR than the Thematic Mapper (TM) in Landsat 4 and
5.
SNR may vary depending not only on sensor characteristics
but also on the signal strength; summer images will have a
higher
SNR than winter images for the same time and place.
While the advantages of high
SNR are well documented in do-
mains such as mineralogical mapping (Chabrillat et al., 2002;
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
July 2004
813
R.V. Platt was formerly with the Department of Geography,
UCB 260, University of Colorado at Boulder, Boulder, CO
80309. He is presently with the Department of Environmental
Studies, Gettysburg College, Gettysburg, PA 17325
(rplatt@gettysburg.edu).
A.F.H. Goetz is with the Center for the Study of Earth from
Space/CIRES/Department of Geological Sciences, UCB 216,
University of Colorado, Boulder, CO 80309
(goetz@cses.colorado.edu).
Photogrammetric Engineering & Remote Sensing
Vol. 70, No. 7, July 2004, pp. 813–819.
0099-1112/04/7007–0813/$3.00/0
© 2004 American Society for Photogrammetry
and Remote Sensing
99-018.qxd 6/9/04 10:51 AM Page 813
Smailbegovic et al., 2000), they have not been thoroughly as-
sessed for land use mapping. It is likely that the influence of
SNR on classification accuracy depends heavily on the classes
of interest. For example, distinguishing irrigated urban land
from irrigated cropland may require a higher
SNR than would
be needed to distinguish spectrally disparate land uses such
as residential land and fallow land.
Finally, the number of spectral bands may influence accu-
racy of land use classification. One study showed the benefits
of increasing the number of bands in classification of the
urban fringe. The study used
SPOT XS data to map farmland
and urban land uses in New Zealand (Gao and Skillcorn,
1998). In this case, using multispectral imagery improved the
delineation of urban areas and farmland because vegetative
land covers were easier to discriminate with a near-infrared
band. In cases where different land uses have similar but sep-
arable spectra, increasing the number of spectral bands will
likely improve mapping accuracy. When land uses are either
spectrally inseparable or clearly distinct, however, additional
bands may not improve classification accuracy. In these cases
the extra bands could add noise and spectral heterogeneity,
resulting in lower classification accuracy.
These studies show that decreasing
GDR, increasing SNR,
and increasing the number of bands may improve classifica-
tion accuracy for land use mapping, but the net benefits often
depend on the particular scene and classification system. In
this study
AVIRIS data was compared with synthetic Landsat
ETM+, fixed at 20 meter spatial resolution to determine the
possible effects of increased number of bands and higher
SNR
for land use mapping at the urban fringe in Colorado.
Image Processing
An AVIRIS flight line was acquired for 30 September 1999
along the northern Front Range of Colorado. A single image
cube was extracted that encompassed much of Fort Collins
along with the surrounding agricultural land and Horsetooth
Reservoir (Figure 1).
In order to convert at-sensor radiance into surface re-
flectance, an atmospheric correction was performed with
High-Accuracy Atmosphere Correction for Hyperspectral Data
(
HATCH). Using spectral features within the data, HATCH creates
pixel-by-pixel estimates of atmospheric composition.
HATCH
814
July 2004
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
TABLE 1. SENSOR CHARACTERISTICS
Landsat
AVIRIS TM/ETM+ SPOT XS
Platform Airborne Spaceborne Spaceborne
Ground Sampling Distance 20 m 30 m 20 m
Number of Bands 224 6 3
(excluding thermal)
Signal-to-Noise Ratio High Moderate Moderate
Launch 1987 1982 1986
Figure 1. Synthetic Landsat ETM+ Band 3 image of Fort Collins and surroundings.
99-018.qxd 6/9/04 10:51 AM Page 814
takes advantage of recent advancements in atmospheric radia-
tive transfer, resulting in highly accurate atmospheric correc-
tions (Qu et al., 2000; Qu et al., 2002). After conducting the
atmospheric correction, bands in the major water absorption
features at 1.4 m and 1.9 m were removed.
In this study, an
AVIRIS image was compared to synthetic
Landsat
ETM+ image derived from AVIRIS. This method elimi-
nated several sources of error that would be present if a real
Landsat image were used. First,
AVIRIS images from 1999 and
earlier contain unsystematic distortions introduced by the
pitch, yaw, and roll of the aircraft (currently a three-axis gyro-
scope is attached to the sensor and records these movements
so that the distortions may later be removed from the images).
As a result, some
AVIRIS images may be difficult to register to
other images with high precision. Secondly, the
GSD of AVIRIS
(20 meters) is finer than that of Landsat (30 meters), necessi-
tating a resampling procedure that would degrade and possi-
bly introduce additional distortions to the image. Finally, the
two images would be recorded at different times of the day, on
different days, with different atmospheric conditions that
would need to be corrected. Though it is likely that the cumu-
lative effects of these differences would be small, they would
no doubt introduce errors into the comparison.
A solution to all of these issues was to create an image
from
AVIRIS that closely matches the output of Landsat ETM+.
This was done with a two-step process. In the first step, the
appropriate
AVIRIS bands were combined to approximate the
following Landsat bands:
Band 1: 0.45–0.52 m (blue)
Band 2: 0.52–0.60 m (green)
Band 3: 0.63–0.69 m (red)
Band 4: 0.76–0.90 m (near infrared)
Band 5: 1.55–1.75 m (short wave infrared)
Band 7: 2.08–2.35 m (short wave infrared)
To create each synthetic
ETM+ band, 7 to 27 AVIRIS bands
were combined. Since each detector is most sensitive to the
wavelength at the center of the sensor bandwidth, the
AVIRIS
bands that fell in the middle of a Landsat band were weighted
more than those that fell toward the edge of the band, accord-
ing to the Landsat
ETM+ filter response function. Before pro-
ceeding, the dynamic range of the synthetic Landsat images
was degraded from 12 bits to 8 bits to approximate
ETM+.
In the second step, the
SNR of the synthetic ETM+ image
was degraded to approximate the SNR present in actual ETM+. In
1999, when the image was taken,
AVIRIS bands had an SNR as
high as approximately 1000 (figures from Robert O. Green, Jet
Propulsion Laboratory, personal communication). Since noise
is inversely proportional to the square root of the number of
bands, the synthetic
ETM+ has even lower noise than actual
AVIRIS
data and is approximately 28 to 37 times greater per band
than that of
ETM+(Table 2). As a result, AVIRIS may outperform
ETM+ even if spatial and spectral resolution were equalized.
To estimate the noise levels of ETM+, the following model
was used (John Barker, NASA/Goddard Space Flight Center,
personal communication):
SNR DN(a b
*
DN )^.5, (1)
where DN is the digital number of a pixel, and a and b are
coefficients for each band calibrated on
ETM+ data from 06 Sep-
tember 2002. The model produced estimated per-pixel
ETM+
SNR (scene averages shown in Table 2). Dividing the DNsby
the estimated
SNR produced an estimated noise level for each
pixel. Gaussian noise images were then created with a stan-
dard deviation equal to this noise level (over and above that of
AVIRIS) for each pixel of each band. These noise images were
added to each synthetic
ETM+ band to approximate the noise in
the actual
ETM+ sensors. The resulting synthetic ETM+ images
very closely approximated the bands and
SNR of actual Landsat
ETM+, only with a GSD of 20 meters instead of 30 meters.
After creating the synthetic Landsat image, a Maximum
Noise Fraction (
MNF) transform was performed on the AVIRIS
cube and synthetic Landsat images to reduce processing time
and noise, (Green et al., 1988. Note:
MNF is referred to as
“Minimum Noise Fraction” in Environment for Visualizing
Images (
ENVI) image processing software). A MNF transform,
similar to a principal components transform, derives a series
of uncorrelated bands and segregates noise in the data. Unlike
a principal components transform, a
MNF transform equalizes
the noise across bands so that image data with variance lower
than noise is not hidden in higher bands. All
MNF bands with
an eigenvalue of less than two were eliminated since these
bands contain mostly noise. The number of remaining bands
equals the dimensionality of the image. In this case, the syn-
thetic
ETM+ data had a dimensionality of five, and the AVIRIS
data had a dimensionality of 30. All subsequent analysis was
conducted on these two data sets.
Figures 2 and 3 show the band loadings for
MNF bands 1
to 5 and 16 to 30. There are 210
AVIRIS bands between 0.4 and
2.5 m. The first few
MNF bands (Figure 2) show loadings that
peak in the atmospheric windows and are not single-wave-
length specific. In fact, the peak loadings fall approximately in
line with the Landsat bands. On the other hand, in the higher
MNF bands (Figure 3) the loadings are much more wavelength
specific as evidenced by the sharp changes throughout the
spectrum. Some of these significant loadings are associated
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
July 2004
815
T
ABLE 2. SIGNAL TO NOISE RATIO (SNR) OF SYNTHETIC AND TRUE LANDSAT ETM+
Average
SNR for
AVIRIS
bands in SNR of Estimated Ratio of
ETM+ synthetic synthetic SNR of synthetic to
Band ETM+ ETM+ true ETM+ true SNR
1 912 2412 87 28
2 1033 2923 96 31
3 982 2778 75 37
4 821 3178 147 22
5 584 2611 102 26
7 377 1958 67 29
Figure 2. Band loadings for MNF 1-5. Note that peak load-
ings of the AVIRIS data roughly correspond to Landsat
bands.
99-018.qxd 6/9/04 10:51 AM Page 815