Relative Radiometric Normalization Performance for Change

Relative Radiometric Normalization
Performance for Change Detection from
Multi-Date Satellite Images
Xiaojun Yang and C.P. Lo
Abstract
Relative radiometric normalization (RRNminimizes radiometric differences among images caused by inconsistencies
of acquisition conditions rather than changes in sudace
reflectance. Five methods of RRN have been applied to 1973,
1983, and 1988 Landsat MSS images of the Atlanta area for
evaluating their pedormance in relation to change detection.
These methods include pseudoinvariant features (PIF), radiometric control set (RCS), image regression ( m l , no-change
set determined from scattergrams (NC], and histogram matching
(MM), all requiring the use of a reference-subject image pair.
They were compared in terms of their capability to improve
visual image quality and statistical robustness. The way in
which different RRN methods affect the results of information
extraction in change detection was explored. It was found that
RRN methods which employed a large sample size to relate
targets of subject images to the reference image exhibited a
better overall performance, but tended to reduce the dynamic
range and coefficient of variation of the images, thus undermining the accuracy of image classification. It was also found
that visually and statistically robust RRN methods tended to
substantially reduce the magnitude of spectral differences
which can be linked to meaningful changes in landscapes.
Finally, factors affecting the pedormance of relative radiometric normalization were identified, which include land-use/
land-cover distribution, water-land proportion, topographic
relief, similarity between the subject and reference images, and
sample size.
Introduction
Spectral data acquired by satellite sensors are influenced by a
number of factors, such as atmospheric absorption and scattering, sensor-target-illumination geometry, sensor calibration,
and image data processing procedures, which tend to change
through times (Teillet, 1986).Targets in multi-date scenes are
extremely variable and have been nearly impossible to compare in an automated mode (Kim and Elman, 1990).In order to
detect genuine landscape changes as revealed by changes in
surface reflectance from multi-date satellite images, it is necessary to carry out radiometric correction. Two approaches to
radiometric correction are possible: absolute and relative (Lo
and Yang, 1998). The absolute approach requires the use of
ground measurements at the time of data acquisition for atmospheric correction and sensor calibration. This is not only
costly but also impractical when archival satellite image data
are used for change analysis (Hall et al., 1991). The relative
approach to radiometric correction, known as relative radiometric normalization (RRN),
is preferred because no in situ
atmospheric data at the time of satellite overpasses are
required. This method involves normalizing or rectifying the
intensities or digital numbers (DN) of multi-date images bandby-band to a reference image selected by the analyst. The normalized images would appear as if they were acquired with the
same sensor under similar atmospheric and illumination conditions to those of the reference image.
In connection with the fund funded project ATLANTA
(ATlanta Land-use ANalysis: Temperature and Air-quality)
which necessitates accurately mapping changes in the landuselland-cover of the Atlanta Metropolitan Region, Georgia,
for the past 25 years using Landsat MSS data, the RRN approach
is the reasonable choice. Different RRN methods developed for
radiometric correction of multi-date satellite images were evaluated to identify the right one for this project. The only published work of a systematic comparison was carried out by
Yuan and Elvidge (1996),who applied seven empirical RRN
methods to two Landsat MSS images of the Washington, D.C.
area. Their comparisons were made visually, using a measure
of agreement based on standard error statistics.
Despite the excellent effort of Yuan and Elvidge, the performance of the different RRN methods deserves a more thorough
analysis for the following three reasons. First, it is noted that
the Landsat MSS images of the Washington, D.C. area used by
Yuan and Elvidge contain a high proportion of clear water and
urban area, which should favor methods relying on these
ground targets in determining the transformation coefficients.
In contrast, the Landsat M s s data for the Atlanta region contain
a much smaller fraction of water and urban area but a larger proportion of forest and cropland distributed over a foothill and
mountainous area. Therefore, these RRN methods may perform
differently when applied to images with more diverse geographical features. Understanding the variation of RRN performance with respect to different scene conditions is important
for reinforcing the absolute and comparative utility of these
methods. Second, the effects of alternative radiometric normalization methods upon the outcomes of change detection are
unknown. To bridge this gap, possible impacts of different RRN
methods on image classification and spectral change detection
will be examined. Finally, it is believed that there are factors
that may affect the performance of radiometric normalization
other than the RRN methods themselves. By comparing the performance variations of RRN methods as applied to two reference-subject image pairs with different levels of similarity, this
Photogrammetric Engineering & Remote Sensing
Vol. 66, No. 8, August 2000, pp. 967-980.
0099-1112IOOI6608-967$3.00/0
Department of Geography, University of Georgia, Athens, GA
30602 (yang&ga.edu).
OGRAMMnI;
iNGll
0 2000 American Society for Photogrammetry
and Remote Sensing
August 2000
967
study attempts to uncover these factors and suggest means to
handle them from an operational perspective.
Relatlve Radiometric Normalization (RRN) Methods
In this paper, only those RRN methods using a reference-subject
image pair were evaluated. These methods may be further subdivided into three groups: statistical adjustments, histogram
matching, and linear regression normalization.
The statistical adjustments approach includes several
methods which are based on the linear adjustment of two
images to resemble each other in terms of their dynamic range
(minimum and maximum DN values), statistical mean and
standard deviation, or other possible statistical variables. l h o
methods -the minimum-maximum normalization and the
mean-standard deviation normalization- have been evaluated
by Yuan and Elvidge (1996).Their results showed that these
methods did not perform well.
Histogram matching is a commonly used radiometric
enhancement technique fully integrated in many image processing software packages (Richards, 1986).The transformation
is a non-linear matching of histograms of two images so that
apparent distributions of brightness values in the two images
correspond as closely as possible. The subject image's histogram
must be equalized first to obtain an intermediate histogram,
which is then modified to match the shape of the reference
image's histogram. In essence, histogram match is a process of
determining a lookup table that will convert the histogram of
one image to resemble that of another. It is, therefore, a useful
technique for matching image data of the same scene acquired
at different dates with slightly different sun angles or atmospheric effects.
Linear regression normalization includes diverse methods
developed o 6 r the past decade. One fundamental premise
behind these methods is that the radiance reaching an airborne
or satellite sensor in a given spectral channel can 6e expressed
as a linear function of reflectivity (Schott et al., 1988).For many
sensors, the digital numbers (DN)in each band are a simple linear function of the radiance reaching the sensor. Thus, the
atmospheric and calibration differences between scenes are
linearly related (Schott etal., 1988; Casseles and Garcia, 1989;
Hall et al., 1991).Accordingly,a linear equation can be used to
perform the normalization: i.e.,
where Skisthe digital number (DN)of band k i n image Son date
1, S i is the normalized digital number of band k on date 1, mk is
the slope or gain, and bk is the intercept or offset. Both mt and
bkare iompited through a linear (leak-squares) regressidn performed on two radiometric control sets samnled from the refer-.ence and subject images, respectively, or on the whole scene of
the image pair. The four major empirical methods used to perform RRN in this way are summarized graphically in Figure 1.
Radiometric normalization using image regression (IR)
simply involves relating each pixel of the subject image with
that in the reference image band by band to produce a linear
equation either through a least-squares regression (Jensen,
1983; Singh, 1989) or a robust regression (Olsson, 1993). The
regression technique accounts for the differences in the mean
and variance between radiance values for different dates. The
calculation of transformation coefficients through a leastsquares regression is illustrated in Figure 1A. This method uses
all the pixels in the reference-subject image pair.
Radiometric normalization using pseudoinvariant features (PIF) was developed by Schott et al. (1988) and Salvaggio
(1993).Pseudoinvariant features are objects with nearly invariant reflectivity from one image scene to another. According to
Schott etal. 11988), these are typically man-made objects whose
-
968
August 2000
reflectance is independent of seasonal or biological cycles. Differences in the brightness distributions of these invariant elements are assumed to be a linear function. Schott et al. (1988)
isolated the urban features from the satellite images using an
infrared-to-red ratio image, which is quite effective in differentiating urban and water from vegetation. Once the urban features pixels are isolated, linear regression equations are
developed to relate the subject images to the reference image
band by band (Figure 1B).This method was used by Henebry
and Su (1993) to rectify radiometrically eleven TM images of
tallgrass prairie span&ng a period from 1987 to 1988. Radiometric normalization usine the radiometric control
sets (RCS) was developed by Hall et ax (1991).The underlying
assumption is that an image always contains at least some pixels that have the same average surface reflectance among
images acquired at different dates. This relationship can be best
examined by comparing the band-to-band scattergrams in
which the pixels along the diagonal line should have little or no
variation through the time period represented by the dates of
the two images. These objects serve as a so-called radiometric
control set (RCS).Instead of taking a single sample set of bright
pixels representing the urban built-up land, this method uses
the two non-vegetated extremes of the Kauth-Thomas (KT)
greenness-brighess scattergram which is constructed using
the first two bands of a Tasseled C ~ transformation.
D
The KT
scattergram isolates a dark control iet, theoretically corresponding to pixels of deep water (e.g., reservoirs) and a bright
control set, representing similar elements of pseudoinvariant
features defined by Schott et al. (1988).The transformation
coefficients are calculated from the individual band means (in
raw digital counts) of the radiometric control sets in each
image (Figure 1C).This method performed well for the visible
and near-infrared bands of a pair of Landsat 5 TM images,
adjusting surface reflectance for the effects of relative atmospheric differences to within 1percent (Hall et al., 1991).
Franklin and Giles (1995) used this method to normalize four
different dates of Landsat MSS images for classification purpose in the Wood Buffalo National Park Remote Sensing
Project.
Radiometric normalization through no-change set determinedfrom scattergram (NC) was developed by Evidge et al.
(1995).The no-change set is obtained from aregion identified
as no change in the scattergram between a subject image and
the reference image band by band. A no-change pixel set contains pixels occupying the core of the water and land data clusters observable in the scattergrams of the near-infrared bands.
The centers of the water and land data clusters are located with
reference to the local maxima for the near-infrared band scattergrams. The water data cluster center is located at low digital
number values near the lower left corner of the scattergrams
close to the origin while the center of the land data cluster is
located near the center of the scattergram (Figure ID). The nochange region for a particular band is determined by linking
the centers of water and land for that band and expanding perpendicularly to the water-land line up to some digital numbers
on either side of the line. Pixels falling within the no-change
region will be used to compute regression lines for all bands.
This method works best if the majority of the pixels of the subject scene at one date have the same land cover and vegetation
growth stage as the reference scene at another date (Evidge et
al., 1995).
Data and Methodology
This study specifically focuses on the RKN methods which have
a solid theoretical background and the potential to be operational. Therefore, only five methods, including the histogram
matching method (Richards, 1986) and the four linear regression normalization methods mentioned above, will be
evaluated.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
h
E
W
A Image Regrerslon (IR) Method
Maans:
3
Reference image - R,
Subject image - S,
i
Variances:
Reference image v
,,
Subject image v,,
(3
6
3
W
Pseudoinvariant Set (PIS)
C
3
-
VRkSk
3
Transformation coefficients:
3
2
Slope: m, = v, I v,,
Intercept: b, = R, m,
k band
2
k
s
c
12
n
--
-
-
-
s
*s,
Transfotmatlon coefficients:
Slope: m, = a, I a,
Intercept: bk= R, m, *Sk
k band
C. Radiometric Control Set (RCS) Method
W
-
-
12
n
DIGITAL VALUE OF REFERENCE IMAGE (R)
DIGITAL VALUE OF REFERENCE IMAGE (R)
E
-
-
LL
0
.
:$.:
PIS of reference image a,
PIS of subject image a,
Meanrr:
PIS of reference image R,
PIS of subject image S,
2
3
V)
w
B. Pseudoinvariant Featurn (PIF) Method
E
D. No Change (NC) Set Determined
from Scattergram
c
(3
Rs(ersnce image
Rw=f--S,
I-
3
3
2
3
V)
V)
-R,
Fi%&imga-vsubjedlmsge-v-
"-
-
No change
,
/
,
pC"7"
Transformation coefficients:
Slope: Q=(&,-D&I(B,-D.J
Intercept: 4, = (D,
-
E& D,
-
b)1(0- D.J
Transformation coemcients:
-
Slop: m.=v,/v,
12
a
'
W~
/
e n t water
v
k-bald
4=Rt - Q %
0
DIGITAL VALUE OF REFERENCE IMAGE (R)
DIGITAL VALUE OF REFERENCE IMAGE (R)
Figure 1. A graphical summary of four major empirical methods of relative radiometric normalization by developing
linear regression equations. (A) lmage regression (IR) method with the solutions based on least-squares transformation (Jensen, 1983; Yuan and Elvidge, 1996). (B) Pseudoinvariant feature (PIF) method (Schott et a/., 1988).
(C) Radiometric control set (RcS)method (Hall eta/., 1991). (D) No-change (Nc) set determined from scattergram
method (Elvidge et a/., 1995). HPW = half perpendicular width. Hvw = half vertical width.
Data and Study Site
Landsat MSS data of the Atlanta Metropolitan area for 1973,
1983, and 1988 were used in this research. Physiographically,
the Atlanta image area mainly covers the foothills of the southern Appalachians in northern Georgia at elevations from 300 to
350 m above mean sea level. The northwestern region (approximately 25 percent of the total area) is the Appalachian mountains. The major land-uselland-cover types are forestland (50 to
60 percent), grassland/cropland (5 to 10 percent), urban builtup land (15 to 25 percent, with 2 to 5 percent of core urban use),
cultivatedlexposed land (3 to 6 percent), and water (2 percent,
mostly deep reservoirs).
The three raw images are illustrated in Plate 1.These
scenes are very different from each other even though they
were processed identically. The 1988 image was selected as the
reference image because of its excellent visual quality when
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
displayed and, most important of all, the availability of largescale aerial photographs for 1988, which allows targets
extracted from the images to be checked. Thus, two referencesubject image pairs were formed, including 1988-1973 and
1988-1983. The three images were acquired by different Landsat satellite sensors. The 1973 image was acquired by the first
generation of Landsat (Landsat-1). Both the 1983 and 1988
scenes were acquired by the more advanced secondgeneration
of Landsat (Landsat-4 and -5). The radiometric differences
among these images are caused by such factors as seasonal
changes, sensor variations, atmospheric properties, and sensortarget-illumination geometry. The relative magnitude of the
collective effects caused by these factors can be perceived by
comparing two raw image pairs. The relative differences
between the 1973 and 1988 images are much larger than those
between the 1983 and 1988 images. These differences can be
August ZOO0
969
-
IMAGE PAIR: 1988 1973
1
Referema l m a p (May 14,1988)
Subject bnage (April 13,1973)
NO Chang. Set R.01.rulon (NC)
Hktogrmw@h(W)
PsaudokmrlantFwtuna (PF)
'
1
hWR.(lw(IR)
R~llom.tfkControl Sob (RCS)
-
IMAGE PAIR: 1988 1983
I
ReferenceImage (May 14,1988)
SubJoctImage (May 9,1983)
Plate 1. Two image pairs: 1988 (reference)-1973 (subject) and 1988 (reference)-1983 (subject). The original
images and those normalized by different methods are shown. They were displayed as a standard falsecolor
composite (RGB-B421) after applying an identical linear stretch. The image dimension is approximately 140 km by
150 km.
appreciated both visually (Plate 1)and statistically (Figure 2).
For the 1973-1988 image pair, the 1973 image has a much
duller appearance, particularly for the forest lands which
should be red or reddish in a standard false color composite
(Plate 1).It has a much narrower dynamic range which can be
observed from the scattergrams (Figure 2). Given these observations, the two image pairs can be further distinguished as an
image pair with larger differences in terms of scene characteristics (the 1973-1988 image pair) and an image pair with
smaller differences (the 1983-1988 image pair). They are,
therefore, ideal for use in evaluating the performance variations
of the RRN methods when applied to diverse data sets. The scattergrams relating each band of the subject images to the reference image were produced, not only for comparing the
dynamic data range as described above, but also to help in
determining the transformation coefficients for the NC method.
Methods
.
-
The experimental procedures include data preprocessing,
selection of sample sets, calculation of transformation coefficients, transformation operation, and performance assessment.
These procedures were used for the four RRN methods requiring the computation of a linear regression equation.
(1) Preprocessing. The georeferencingstrategy used here
was to geometrically correct one image, i.e., the 1988 MSS
image. Then the corrected 1988 scene was used as the reference
image for image-to-imageregistration. The resultant planimetric accuracy was excellent (20.5 pixel size).
(2)Selecting Sample Sets. With the exception of the image
regression (E)method which makes use of the whole image
scene, the other three methods employ very different strategies
to select sample sets. The theoretical background of these strategies has already been explained.
Pseudoinvariant features (PIFS)were segmented by intersecting the mask from the infrared-to-red ratio image (Band 41
Band2) with the mask from the infrared image (Band 4). These
masks were isolated using different threshold values obtained
by interactive observations (Table 1).The intersect operation
corresponds to an "and" logical operation: i.e.,
PU
=
{(=Band
4
5
tl) and (Band 4 2 t,)
I
where t, and t2are the threshold values. The resultant PU sample sets are mainly the scene elements covering downtown
Atlanta, the Hartsfield International Airport, and large commercial corridors along the Interstate Highways (I75 and 185).
The initial radiometric control sets (RcS) were segmented
by intersecting the mask from the brightness band with the
mask from the greenness band. The two bands were computed
using the following equations (Crist and Kauth, 1986):
Brightness = 0.3320 * (Band 1)+ 0.6030 * (Band 2)
+ 0.6750 * (Band 3) + 0.2620 * (Band 4)
(3)
-
IMAGE PAIR: 1988 (REFERENCE) 1973 (SUBJECT)
A. Band 1 (0.5- 0.6 pm)
-
r
m
B. Bmd 2 (0.6 - 0.7pm)
C. Bmd 3 (0.7- 0.8 pm)
-
m
N
2
5, g - 2 '..:
- -.
v
E?
rn
Y
m
u
E
u
5
::
o+.,.
I
-
7
p .Zi
N
0
:y:=:;:.-:
a*
::
'.'.
p&Y
+'
96
128
Band 1 (1 873) - CN
Band 2 (1873)
- CN
D. Bmd 4 (0.8- 1 1 pm)
r
I
./
E
I?
P
m
:;
0
64
-
N
Band 3 (1973) -
sa
CN
ria
Band 4 (1973) - DN
-
IMAGE PAIR: 1988 (REFERENCE) 1983 (SUBJECT)
-
E. Band 1 (0.5- 0.6 pm)
F. Band 2 (0.6 -0.7vm)
G. B a d 3 (0.7- 0.8 pm)
Band 2 (1983) -ON
Band 3 (1983) ON
w
H. Bmd4(0.8- 1.1pm)
R
6:
-
8
m
m
-'o
v
u
5
m
R
0
-
Band 1 (1983) DN
-
Band 4 (1983) - DN
Figure 2. Scene scattergrams of 1988 (reference image) versus 1973/1983 (subject image) digital number values (DN) band
by band. Note that the regions having the densest numbers of pixels are displayed in white, flanked by zones of gray and
black, corresponding to declining numbers of pixels. Areas outside or intercepting the data clusters, which have no pixels
present, are white.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August ZOO0
971
SETSWITH THE THRESHOLD
VALUESDETERMINED
TABLE1. THEPIF SAMPLE
THROUGH INTERACTIVE
OBSERVATIONS
Image
t1
PIF sample size
tz
*The threshold value [70) was used to exclude the cloud pixels in the
1983 image. Any pixels with digital value larger than 70 were considered to be cloud pixels, and should be isolated.
Greenness
=
(-0.2830) * (Band 1)+ (-0.6600) * (Band 2)
(3) Calculating Ti-ansformation Coefficients.Once the sample sets were defined, the transformation coefficients can be
computed using the equations provided for each method as
shown in Figure 1.Table 4 shows the slopes and intercepts of
the linear regression in the form of Equation 1.
( 4 ) Ti-ansformation Operations. Using the transformation
coefficients obtained above for each method, each subject MSS
image is then radiometrically corrected by applying Equation 1.
(5) Peqformance Assessment. Four types of evaluation criteria were used in comparing the performance of each RNN
methods. They are discussed in more detail in the next section.
Results and Discussions
(4) Visual Claseness and StatLstlc Robustness
Comparing the visual appearance of normalized products is the
most straight forward way to judge the overall performance of
These masks were isolated using different threshold values
these methods. In doing so, both the radiometrically normalobtained by interactive observations. The equations for segized images and the reference image are displayed side by side
menting the initial radiometric control sets (RCS)are
on the monitor screen, and the visual closeness of each normalized image to the reference image is determined qualitatively.
Initial dark sets = {(greenness5 tl) and
If the visual difference between the normalized image and the
(brightness 5 tz))
subject image is small, the subject image can be regarded as
radiometrically adjusted to the reference image. However,
Initial bright sets = {(greenness5 tl) and
visual comparison is prone to subjectivity. Therefore, a quanti(6) tative comparison is also needed.
(brightness r tz)}
The root-mean-square error (RMSE) is used to measure the
where tl and t, are the threshold values (Table 2). The final dark statistical agreement of a normalized image with the reference
set was defined by selecting the pixels from the lower half of
image, as fdlows:
the initial set because the upper half could contain dark pixels
contaminated by sun glint or sediments. This set mainly consists of scene elements of large reservoirs. The final bright set
was defined in a similar way, but the pixel values were applied
in reverse order to exclude vegetated pixels. This set mainly
where Si, is the radiometrically normalized digital number of
consists of scene elements covering downtown Atlanta and the
band k in the image S (subject image) on date 1, Rkis the digital
Hartsfield International Airport.
number of band k i n image R (reference image) on date 2, and
The no-change (NC)sets were defined using the following
lscenel is the total number of pixels of the scene. Thus, the digiequation (Evidge et al., 1995):
tal numbers of pixels of the radiometrically normalized image
are compared with those of the reference image of the same
band. If the difference between these numbers is small, the
RMSE will be small, implying that the subject image is radiometrically more similar to the reference image.
where m,,, b3,, m4,, and b,, are initial estimates form,, b,, m,,
To facilitate the visual comparison, a high quality hardand b, through locating the centers of water- and land-surface
copy in color was produced (Plate 1). Images on these prints
data clusters from Band 3 and Band 4 scattergrams; and HVW,
were displayed as a standard false color composite after
and H V are
~ the corresponding half vertical widths of the noapplying an identical linear stretch to each image. For the
change regions in the scattergrams. The initial estimates for
1983-1988 image pair, it was found that the images produced
m,,, b30r m,,, and b,, were computed using the equations given
by histogram matching (HM) and image regression (IR)were
by Evidge et al. (1995,p. 1259).The HVW and HPW has a simple
more alike, and gave better results than those generated by
relationship
other methods. They were followed by images generated by nochange set (NC) determined from the scattergram method, radiometric control set (RCS)method, and pseudoinvariant feature
(PIF) method, in this order. The images produced by the HM,IR,
where m is the slope of the initially estimated axis for a given
NC,and RCS methods exhibited much improved visual closeband, HPW is the half perpendicular width which is assigned as
ness to the reference image than that of the 1983 raw image.
ten digital numbers.
Only the image produced by the PIF method gave the worst
+ 0.5770 * (Band 3) + 0.3880 * (Band 4)
TABLE2. THERADIOMETRICCONTROL
SETS(RCS)WITH THE THRESHOLD
VALUESDETERMINED
THROUGH
~NTERACT~VE
OBSERVAT~ONS
Bright Control Set
Image
tl
tz
Initial Size
Final Size
tl
tz
Initial Size
Final Size
1973
1983
1988
77
77
93
41
52
27
47,652
61,250
64,925
25,992
37,056
39,813
100
77
114
100
118-163*
106
44,764
11,025
20,825
23,823
7,044
11,025
*The upper limit of brightness (163) was used to exclude the cloud pixels in the 1983 image. Any pixels with a brightness value larger than
163 were considered to be cloud pixels, and should be isolated. Note that both the brightness and greenness were converted into an &bit
unsigned data format.
972
August 2000
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
TABLE3. THE NO-CHANGE
(NC) SETSWITH PARAMETERS
DEFINED
Image
Pair
1973-1988
1983-1988
Band 3
m30
1.864
1.078
Band 4
b30
mW3
-8.840
-9.057
10
10
~
w
3
21.15
14.704
m40
b40
ww4
H W 4
Size of NC
Set
2.081
1.265
-6.540
-8.874
10
10
23.09
16.13
4,485,064*
5,364,275**
* or 67.24% of the total scene pixels. * * or 80.43% of the total scene pixels.
result because its overall visual appearance was very different
from that of the reference image. As for the 1973-1988 image
pair, it seemed that none of the five methods had improved substantially the overall closeness of their resultant products to
the reference image when compared to the raw image. However, a close examination revealed the following descending
order of improvement in visual quality: the HM, IR, PIF, NC, and
RCS methods.
Obviously, radiometric normalization of the 1983 image
(subject) to the 1988 image (reference) is better than that
between 1973 and 1988. This performance variation may stem
largely from the scene variations in the two image pairs. Variations in vegetation growth season and land uselland cover are
more drastic in the 1973-1988 image pair than those in the
1983-1988 image pair. Also, there is a difference in the quality
of the sensors between Landsat-1 and Landsat-5 for the 1973
and 1988 MSS images (see Thome et al., 1 9 9 7 ) . Clearly, radiometric normalization performs better if the subject and reference images exhibit similar characteristics in vegetation
growth season (as affected by temperature and rainfall), the
spatial distribution of land uselland cover, and similarities in
sensor characteristics.
The RMSES computed for each band for the five methods of
radiometric correction are shown in Table 5 and Figure 3 ,
which also show the number of pixels used by each method to
determine the transformation coefficients for the linear regression equations. For comparison purpose, RMSEs of the raw subject image data without any radiometric normalization were
also computed for all the four bands.
If RMSES are used to evaluate the different methods of radiometric normalization, it is clear that the IR, HM, and Nc methods
are better than the RCS and PIF methods. For the 1983-1988
image pair, the IR method is the best with the smallest average
RMSE,followed by the Nc, HM, KCS, and PIF methods. For the
1973-1988 image pair, the IR method is also the best, followed
by the HM, N C , PIF, and R ~ methods.
S
All the methods have
reduced to varying degrees the radiometric difference between
the subject image and the reference image after radiometric normalization, as indicated by the fact that the average RMSEs of
each normalized image are much smaller than those for the raw
image data. It is also clear that radiometric normalization of the
1983 image to the 1988 image is statistically better than that
from the 1973 image to the 1988 image as shown by the fact that
the average RMSEs of the normalized images of 1983 are generally smaller than those of 1973. In this respect, the statistical
comparison further confirms the visual one.
The statistical comparison also helps to pinpoint the much
finer differences among the different methods. For example, by
comparing the RMsEs band by band, it was revealed that each
method performed best for particular spectral bands. The visible wavebands, i.e., Band 1(0.5 to 0.6 p m green) and Band 2 (0.6
to 0.7 p m red), can be better normalized radiometrically than
the two near-infrared bands (Band 3: 0.7 to 0.8 pm, and Band 4:
0.8 to 1.1 pm). Such an observation was also made by Yuan and
Elvidge ( 1 9 9 6 ) .The reason behind such a variation is not clear.
However, this poorer statistical agreement for the two infrared
bands may be related to larger variations in vegetation growing
conditions (such as season and moisture) between the two
images, which are better detected in the infrared portion of the
electromagnetic spectrum.
Another important point to note from Table 5 is that the
sample size of targets has affected the results of radiometric
normalization. The image regression (IR)method and histogram
matching (HM) make use of all pixels in the image scene. The
no-change ( N C ) set method also employs over 50 percent of the
pixels in the scene. All of these three methods are better overall
performers in the evaluation. Both the radiometric control set
(RCS) and the pseudoinvariant feature (PIF) methods use a small
percent of pixels in the image scene, and their performance suffers because of this feature (Table 5).
Impacts of RRN Methods on Image Frequency Distribution
Radiometric normalization alters spectral signatures, contrasts
between (feature) categories, or variances and covariances of
TABLE4. RADIOMETRICNORMALIZEDCOEFFICIENTS
(SLOPE-m
A N D INTERCEPT-b)
A. 1973-1988
Band
Method
PIF
RCS
IR
NC
B.
Band 2
1
Band 3
117,
b,
In2
b2
m3
1.275
1.219
0.442
0.379
-28.495
-30.301
-1.877
-0.334
1.119
1.133
0.425
0.272
-12.501
15.367
2.776
4.610
1.135
1.063
0.273
0.932
b2
m,
Band 4
b3
-1.697
4.661
42.791
21.527
b4
m4
1.369
1.110
0.350
1.268
-5.962
0.337
48.621
18.675
1983-1988
Band
Method
m1
Band 2
1
b,
m2
Band 3
Band 4
b3
m4
b4
-
PIF
RCS
IR
NC
1.598
1.114
0.551
0.652
34.937
-15.375
-2.803
-5.400
1.228
0.648
0.246
0.290
Note: PIF-Pseudoinvariant
Feature Method; RCS-Radiometric
determined from Scattergrams method.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
-83.740
14.925
0.400
-1.929
1.621
1.210
0.508
0.775
Control Set Method; IR-Image
-51.631
-9.037
22.987
7.932
2.153
1.440
0.655
1.010
-53.067
-3.948
27.668
10.177
Regression Method; and NC-No-change set
August 2000
973
TABLE5. STAT~STICS
OF SAMPLING
SETSAND RMS (ROOT-MEAN-SQUARE)
ERRORS
A. 1973-1988
Sample Sets in 1988
Image
Sample Sets in
1973 Image
Method
Number
Raw
PIF
6669836
173280
25992(D)
23826[B)
49818(T)
6669836
4485064
6669836
RCS
IR
NC
HM
%
100
2.60
0.39
0.36
0.75
100
67.24
100
Number
6669836
171500
39813(D)
11025(B)
50838(T)
6669836
4485064
6669836
Root-Mean-Square (RMS) Errors
%
Band 1
Band 2
Band 3
Band 4
Average
100
2.57
0.60
0.17
0.78
100
67.24
100
18.606
5.457
14.038
8.623
17.734
16.031
25.906
20.565
19.069
12.669
7.265
8.651
20.029
22.682
14.657
3.916
3.705
4.200
7.680
7.728
8.630
8.686
11.833
10.230
11.016
15.460
13.357
7.825
9.682
9.133
B. 1983-1988
Sample Sets in
1983 Image
Method
Number
Raw
PIF
6669836
170275
44100(T)
6669836
5364275
6669836
- --
Sample Sets in 1988
Image
Root-Mean-Square (RMS) Errors
Number
Yo
Band 1
Band 2
Band 3
Band 4
Average
100
2.55
6669836
171500
100
2.57
14.355
6.547
41.493
14.159
11.004
15.084
13.545
16.085
20.099
12.976
0.66
100
80.43
100
50838(T)
6669836
5364275
6669836
0.78
100
80.43
100
3.400
3.354
3.505
6.398
6.658
7.380
7.310
7.780
8.056
9.529
10.020
10.473
6.659
6.953
7.354
O/o
RCS
IR
NC
HM
Notes. For the Radiometric Control Set (RCS) method, we provide the average numbers of the sample sets; D-the
bright control set, and T-the total number combining the two sets.
spectral bands, which may further affect the results of image
classification and change detection. The way in which different RRN methods alter statistical measures of image data is a
fundamental concern for remote sensing analysts in selecting
the best method for specific applications. To explore this relationship, the frequency distribution of each normalized image
was examined because it can affect the separatibility of spectral
classes in automatic classification. Images with disperse distributions are easier to be clustered than those with compact distributions although this relationship may be complicated by
the existence of other statistical phenomena (Arbia et al., 1996).
There are a number of different measures which can be used to
evaluate the spread of dispersion of a distribution, such as
range, quartile deviation, mean deviation, standard deviation,
variance, and coefficient of variation (Burt and Barber, 1996).
Based on the understanding that most (raw and normalized)
images in this study have shown a single-modal distribution of
frequency, two measures, namely dynamic range and coefficient of variation, were chosen for comparing the dispersion of
each image band by band generated by different RRN methods.
Dynamic range, or the difference between the maximum
and the minimum digital numbers, is an important feature
related to the information content. Data with larger range often
contain much finer details that are very helpful in feature identification and information extraction. The alteration of range
from a larger one to a smaller one risks a loss of information.
Coefficient of variation (CV) or the relative variability of a
frequency distribution is measured by the ratio of the standard
deviation (a)
to the mean (p):i.e.,
Although standard deviation and variance can also be
used to judge the dispersion, it is improper to rely on them in
comparing multiple distributions with different means (Burt
and Barber, 1996),for which coefficient of variation is a better
974
Auguct 2000
dark control set, B-the
measure. Obviously, the larger the coefficient of variation, the
more dispersed is the distribution.
The dynamic range and coefficient of variation computed
for each band for different RRN methods are shown in Figure 4,
which also shows the averages of these measures for four bands
as well as the measures for the NDVI images. For comparison
purpose, these same measures for the raw images were also
computed.
For the 1973 scene, the image produced by the HM method
exhibits the largest dynamic range as a whole, followed by the
PF, RCS, NC, and IR methods (Figure 4E). The 1973 NDVI image
computed from the normalized image by the PIF method has
the largest dynamic range, followed by the RCS,HM,NC, and IR
methods (Figure 4F). For the 1983 scene, the image produced
by the PF method has the largest dynamic range, followed by
the RCS, HM,NC, and IR methods (Figure 4K). The 1983 N D X
image computed from the image normalized by the PIF method
is also the best in terms of its dynamic ranges, followed by the
RCS, HM,IR, and NC methods (Figure4L). It should be noted that,
in both the 1973 and 1983 scenes, the images normalized by
the PF and RCS methods as well as their NDVI images have consistently larger dynamic range than their raw images (Figures
4E, 4F, 4K, and 4L).In terms of dynamic range, the outstanding
status of the PF and RCS methods over all methods tested is
consistent for the 1983 images band by band but variable for the
1973 images on a band basis.
As for the coefficient of variation (CV), the PIF and RCF
methods are consistently the best among all the five methods
for both the normalized images and the NDVIimages as a whole
in two scenes (Figures4E, 4F, 4K, and 4L).
A very important point to note from these comparisons is
that those RRN methods which perform best in the visual and
statistical comparisons (such as the IR and NC methods) tend to
reduce the dynamic ranges and coefficient of variations of the
raw images. These methods must be used with caution because
any drop in dynamic range and coefficient of variation will
reduce the dispersion of frequency distribution of an image
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
RRNMuhods
C.
1073, Eand?
1 1
RRN Math&
D.
lOTS.Band4
I I
RRNM*hob
€ 1873. Avemga dfour bands
Figure 3. Comparisons of Root-Mean-Square Error (RMSE) for the raw images and their normalized products by different RRN methods. The verical axis is RMSE.
Raw, raw image; NC, nochange set determined from scattergram method; RCS, radiometric control set method; HM, histogram matching; PIF, pseudoinvariant
feature method; and IR, image regression method with the solutions based on least-squares transformation.
A. 1873,Band 1
B.
D.
1973, Band 4
E. 1973, Avenge of four bands
0.
1983, & d I
H. 19113, Band 2
J.
1983. Band 4
K.
1073, Eund 2
1083, Avomar of Mt ban&
C.
1973, Band 3
F. 1973. NWI (Band 4 -Band gl(Band 4 + Band 2)
I.
1083, Band 3
L. 1983, NDVl (Band 4 -Band 2)I(Band 4 + Band 2)
Figure 4. Comparisons of dynamic range and coefficient of variation (CV) for the raw data and their products normalized
by different methods. Note that in all except F and L, the CV is multiplied by 100. Raw, raw image; NC, no-change set
determined from scattergram method; RCS, radiometric control set method; HM, histogram matching; PIF, pseudoinvariant
feature method; and IR, image regression method with the solutions based on least-squares transformation.
976
August 2000
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
band, which might adversely affectthe separatibility of spectral classes in image interpretation and classification.
Impacts of RRN Methods on Magnitude of Change
Radiometricnormalization minimizes the impacts caused by
acquisition conditions (such as sensor variations, atmospheric
effects, and viewing geometry) other than the effects due to
changes in vegetation growth seasons and land uselland cover.
In other words, regardless of the relative normalization methods employed, the normalized images should still exhibit certain levels of differences that are related to surface reflectance
change. In this sense, an RRN method that dramaticallyreduces
the numeric differencebetween two images may result in suppressing the magnitude of change related to vegetation growth
seasons and land uselland cover. In actual applications, ignoring this aspect may produce incorrect results for change detection. Thus, how different RRN methods will affect the change
magnitude needs to be investigated. To explore this relationship, a simple change detection algorithm, i.e., image subtraction, was applied to the data: i.e.,
where R s is
~ the digital number of band k i n the output image,
Rkis the digital number of band k i n the reference image, and
Skisthe digital number of band kin a normalized image or subject image. Then, the global means of these output images were
compared band by band to determine the magnitude of change.
Clearly, a smaller mean indicates a lower magnitude of change
between two images.
The mean of digital number (DN)for the images produced
by a simple subtraction of a normalized image from the 1988
image is shown in Figure 5. For comparison purposes, the DN
mean for the outputs produced by the subtraction of a subject
image from the reference image was also computed. In Figures
5E and SF, the average of the means of four bands were used as
an absolute value.
Based on Figure 5,it is clear that the difference in digital
number between the 1988 image and the images produced by
the RRN methods with good visual and statistical characteristics
tends to be smaller (Figures 5E and 5K). This difference is
inversely proportional to the RMSE as discussed before. However, these two comparisons are quite different in nature. The
RMSE is used as a measure of goodness of fit in judging the statistical robustness for a specific method. Image differencing is a
standard algorithm used in change detection analysis (Jensen,
1996).
These comparisons revealed that the HM,IR, and NC methods tend to suppress a larger part of change between the two
images. This trend is extremely polarized for the 1988-1983
image pair in which a 96 to 99 percent change in digital number has been suppressed by these three methods. In contrast,
the RCSand PIF methods have suppressed a 30 to 44 percent
change for the 1988-1973image pair and a 62 to 66 percent for
the 1988-1983pair, leaving more room for change-detection
analysis related to phenomena such as land-cover changes.
Thus, these RRN methods which tend to substantially suppress
the magnitude of spectral changes between two images must be
used with discretion in change analysis.
Operational Perspective
Operational perspective is another important concern in selecting an RRN method. Both computation intensity and ease of
implementation of each method can be used to assess this
aspect. Computation intenqity should be measured in terms of
the amount of processing time needed. However, these methods were generally implemented in an interactive mode; thus,
it was impossible to measure the time in an absolute way. Also,
with advances in computation power, the issue of computation
PHOTOGRAMMETRICENGINEERING & REMOTE SENSING
intensity is of limited meaning and, thus, was not considered
here. The assessment of operational perspective focused on the
ease of implementation for each RRN method.
As can be seen from the previous sections, the major bottleneck to implementing a specific method is its procedure in segmenting sample sets. Because both the histogram matching
[HM)and image regression (R)
methods do not require this procedure, they are more easily implemented. Indeed, many image
processing software packages have the HM method as a built-in
function as a radiometric enhancement technique. The other
three methods (PIF,RCS, and NC) are generally more difficult to
implement. The PIF method is easier to be executed than the RCS
method because the PIF method only requires a set of segmented bright pixels while the RCS method needs two sets of
samples. T L CC
~ method is the most difficult to implement
because it involves a much more complicated procedure for
selecting the no-change set.
Factors Affecting Radiometric Normalization
The results of the RNN performance evaluation undertaken in
this research suggest the following considerations in the
choice of an RRN method for change analysis: reference image,
sample set(s),sample size, and scene characteristics.
Refemme Image
In RRN, the role of the reference image is critical because all of
the other original (subject)images are to be adjusted to it. A reference image should have better visual quality than other
images in the same data set as well as a larger dynamic range so
that the information can be maintained as much as possible (Lo
and Yang, 1998).
In selecting a reference image, priority should be given to
those which have good ground reference data available within
a close time span (Jensen et al., 1995).These reference data can
be large-scale aerial photographs or ground observations and
measurements to help in the identification of radiometric control sets used bv different RRN methods. For a laree amount of
image data covering a long time span, it is preferible to select
the scene closest to the middle of the time sequence instead of
the most recent scene as the reference image.-1n doing so, the
difference in land uselland cover between the subject image
and the reference image is minimized, thus maximizing the
similarity between the two images.
Sample Sets
Sample sets are also called radiometric normalization target
sets. They contain the actual scene elements through which
the transformation coefficients are derived. There are three
types of sample sets used by different methods: pseudoinvariant feature (PIF)sets, (dark-bright) radiometric control sets
(RCS),and no-change (NC)sets. The majority of the RRN methods employ a classic thresholding technique through a histogram or a two-dimensional scatterplot to segment these sample
sets. While this technique has certain strengths, it also has a
number of weaknesses. First, this technique is generally not so
easy to implement because it involves many complicated procedures. It also requires intensive computation. These are not
quite desirable from an operational point of view. Second, this
technique only accounts for the targets' statistical behaviors,
but innores other important criteria, such as the targets' dimension~physicalprop&ies, and environmental settings, which
are crucial for ensurine Durer scene elements to be seemented.
According to ~ c & & d tet al. (1990),the radiomgtric normalization targets should be at approximately the same elevation as the other features within the scene so that the atmospheric condition estimated from these targets should be identical to that over other parts of the scene. Additionally, these targets should be in a relatively flat area so that incremental
August 2000
977
-
4
-
0
-
-4
-
18 --
I
I
I
a4
."'
A.
-
1988 1973, Bmd 1
Rm
NC
: RC9
HM
PIF
'
IR
'
-
B.
1 9 1 1013, Band 2
-
-"'~m '
NC
' RCS ' HM
' PIF '
IR
lW8-1973,bndl
C.
0 . a --
038 -
024 -
0.12
-
0
-
.-IRm
D.
-
1988 1873, Band 4
-
0
-
ISM 1B63, Band I
'
IR
'
---
-
-
-10
--
-20
--
-30
--
a'
0.
PIF
'
F. 1n86 -1913, NDVl
(B~d4gud2YIBd4+Bsnd2)
E. 1986 1973, Avrnga of four ban&
10
' NC ' RCS ' HM
H. 1-
6
--
0
-
m
I
I
a-
Rm '
L
-
-
NC
' RCS ' HM ' PIF '
1R
-I2'
Rm
1.
190, b n d 2
NC
RCS '
HM
'
PIF
'
-
IS88 1983. Band 3
I
Rnr '
K. 1-8
-
IS83, A m r g . of*
band8
IR
NC
' RCS ' HM ' PIF '
I
IR
F. 1988 -19B3, NDVl
(Band 4 -Bad 2)I(Bend 4 + Bmd 2)
Figure 5. Comparisons of DN means for the outputs after applying a simple change detection algorithm, namely, image
subtraction. In E and K, the average of means of four bands in absolute value was used. Raw, raw image; NC, nochange set determined from scattergram method; RCS, radiometric control set method; HM, histogram matching; PIF,
pseudoinvariant feature method; and IR, image regression method with the solutions based on least-squares
transformation.
978
A u g u s t 2O/j(l
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
changes in sun angle from date to date will have the same proportional increase or decrease in direct beam sunlight for all
normalization targets.
For defining the dark radiometric control targets, the
dimension of the water body may be important. Targets with
small dimensions may suffer from background contamination
because the apparent reflectance of water targets increases
considerably as a result of a signal component of the total solar
flux reflected by the targets and transmitted to the sensor by
atmospheric scattering (Hill and Sturm, 1991).For defining the
pseudoinvariant features (pw)/brightradiometric control targets, one has to consider not only the statistical behavior of the
samples but also their physical properties, such as the surface
moisture and surface reflectance characteristics. Maior changes
in surface moisture will change the overall reflectivity of thg
features that were assumed to be invariant (Schott et al., 1988).
Features with a non-Lambertian surface, such as a city's downtown, may also cause problem in defining the PIF samples. To
comDensate for these issues. it is recommended that the statist i d y based thresholding method be used jointly with skillbased visual observations. The latter can be used to select those
apparently suitable targets object by object through an onscreen digitizer. In doing so, the selection procedures can be
substantially streamlined while excellent results are being
maintained.
The no-change (NC) set was merely defined through a statistical approach (Evidge et a].,1995). Locating the water center may be tedious when a scene does not have enough water
targets. This may affect the estimation of the initial transformation coefficients and, hence, the sample characteristics. Defining the width of the no-change region is also a tricky issue,
because it affects not only sample size but also sample
characteristics.
Sample Slur
The size of the sample set has a significant impact on the results
of radiometric normalization. A larger sample size often gives
a better overall performance of an RRN method, both visually
and statistically. However, it does not mean that the radiometrically normalized products can be used to yield more accurate
change-detectionresults. In addition, a larger sample set may
have more statistical outliers caused by noise in the database,
which may substantially degrade the performance of radiometric normalization. For these considerations, the RRN methods of
image regression (IR) and histogram matching (HM)should be
used with special care. In particular, the HM method may alter
the relative distribution of the land uselland cover because of
the non-linear nature of its transformation. This will affect the
subsequent image classification and change detection.
Scene Characteristics
Scene characteristics that may have a profound impact on
radiometric normalization include land-uselland-cover distribution, proportion of water and land, and topographic relief.
Land-uselland-cover distribution determines the nature of the
sample set used to derive the transformation coefficients. The
proportion of water and land may affect the effectiveness of
different RRN methods. A higher proportion of these features
means that ideal sample sets with large size and excellent quality can be selected using appropriate techniques. It should
favor those methods relying on these spectral reference targets,
such as the radiometric control set (RCS) and the no-change (NC)
set methods. The research conducted by Yuan and Elvidge
(1996) used a pair of Washington, D.C. Landsat MSS scenes
which contained a very high proportion of clear water and
urban area (35 to 50 percent of the total image area). Their evaluation indicated that the NC and the RCS methods were better
than all the other methods including the image regression (IR)
method, pseudoinvarient feature (PIF)method, and several
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
other lesser methods such as haze correction, minimum-maximum normalization, and mean-standard deviation normalization. When applied to the Landsat Mss scene of Atlanta in our
evaluation, the NC and the RCS methods did not give the best
results. The Atlanta scene has few water features (less than 2
percent), the majority of which are rivers and resewoirs. The
urban use occupies 15 to 25 percent of the total image area, of
which the core urban use is only 2 to 5 percent. Radiometric
normalization generally performs better for a scene covering
flat areas than for hilly areas because scene variations due to
topographic relief (shadows) are more complicated, which
often prevent an effective implementation of an RRN method.
On the whole, radiometric normalization performs well for the
subject and reference images that are similar in terms of scene
characteristics.
Conclusions
The results of the performance evaluation of five different RRN
methods, as applied to the multi-date Landsat MSS data of the
Atlanta region, have shown that all five methods can reduce
radiometric differencesbetween the subject image and the reference image to varying degrees. Those RRN methods using a
large number of sample pixels in deriving the transformation
coefficients for the linear regression equations (such as the IR
and NC methods) were the best overall performers. But they
tended to shrink dynamic range and coefficient of variation of
the raw images, reducing the dispersion of frequency distribution of an image band, which could affect adversely the seDaratibility of spectral classes in image interpretation and classification. In contrast. both the PIF and RCS methods increased the
two measures conHistently for both the normalized images and
the NDVi images as a whole in two scenes. Those visually and
statistically robust RRN methods (such as the HM, IR, and NC
methods) tended to substantially reduce the magnitude of
spectral change which might be related to meaningful changes
in landscapes. In contrast, the RCS and PIF methods cut down
only a moderate degree of spectral change between the two
scenes, thus favoring better change detection.
Radiometric normalization of the 1983-1988 image pair is
better than that of the 1973-1988 image pair, leading to the
observation that the choice of RRN method should be based on
the nature of the images, notably land-uselland-cover distribution, water-land proportion, topographical relief, and similarity between the subject and reference images. Another
important consideration is the ease with which each method
can be implemented in a conventional image processing
platform.
Finally, to ensure the proper implementation of an RRN
method in a practical application, some operational guidelines
for selecting the reference image and defining the sample sets
are given in this paper. A good reference image should display
the best visual quality with the largest dynamic range. Preferably, ground data, atmospheric condition data, and sensor calibration data for the image should be available. The reference
image should not be too far away in time from the subject
image. For defining the sample sets, one should consider not
only the targets' statistical behaviors, but also their dimension,
physical properties, and environmental settings, to ensure that
only pure scene elements are segmented.
-~ -
Acknowledgment
The research reported in this paper has been supported by a
NASA EOS Interdisciplinary Science (IDS) research grant for
Project ATLANTA.
Arbia, G., R. Benedetti, and G. Espa, 1996. Effects of the MAUP on
image classification, Geogmphical Systems,3:123-141.
August 2000
979
Burt, J.E., and G.M. Barber, 1996. Elementary Statistics for Geogmphers, Guiford Press, New York, 640 p.
Crist, E.P., and R.T. Kauth, 1986,The Tasseled Cap de-mystified,Photogrammetric Engineering 6 Remote Sensing, 52(1):81-86.
Eckhardt, D.W., J.P. Verdin, and G.R. Lyford, 1990. Automated update
of an irrigated lands GIs using SPOT HRV imagery, Photogmmmetric Engineering & Remote Sensing, 56111):1515-1522.
Elvidge, C.D., D. Yuan, D.W. Ridgeway, and R.S. Lunetta, 1995. Relative
radiometric normalization of Landsat Multispectral Scanner
(MSS) data using an automatic scattergram-controlledregression,
Photogmmmetric Engineering b Remote Sensing, 61(10):
1255-1260.
Franklin, S,E,, and P.T. Giles, 1995. Radiometric processing of aerial
and satellite remote-sensing imagery, Computers and Geosciences, 21(3):413-423.
Gilabert, M.A., C. Conese, and F.Maselli,1994. An atmospheric correction method for the automatic retrieval of surface reflectances from
TM images, International Journal of Remote Sensing, 15(10):
2065-2086.
Guyot, G., andX.F. Gu, 1994.Effect of radiometric corrections on NDVIdetermined from SPOT-HRV and Landsat-TM data, Remote Sensing of Environment, 49369-180.
Hall, F.G., D.E. Strebel, J.E. Nickeson, and S.J. Goetz, 1991. Radiometric
rectification: Toward a common radiometric response among multidate, multisensor images, Remote Sensing of Environment,
Jensen, J.R., K. Rutchey, M.S. Koch, and S. Nanunalani, 1995. Inland
wetland change detection in the Everglades Water Conservation
Area 2A using a time series of normalized remotely sensed data,
Photogmmmetric Engineering b Remote Sensing, 61(2):199-209.
Kim, H.H., and G.C. Elman, 1990. Normalization of satellite imagery,
International Journal of Remote Sensing, 11(8):1331-1347.
Lo, C.P., and X. Yang, 1998. Some practical considerations of relative
radiometric normalization of multidate Landsat MSS data for land
use change detection, Proceedings of ASPRS/RTI 1998 Annual
Convention, Tampa, Florida, pp. 1184-1193.
Olsson, H., 1993. Regression functions for multitemporal relative calibration of thematic mapper data over Boreal forest, Remote Sensing of Environment, 46:89-102.
Pons, X., and L. Sold-Sugrafies, 1994. A simple radiometric correction
model to improve automatic mapping if vegetation from multispectral satellite data. Remote Sensing
. of. Environment. 48:
191-204.
Richards, J.R., 1986. Remote SensingDigital Image Analysis: An Introduction, Springer-Verlag. Berlin, 281 p.
Salvaggio, C., 1993. Radiometric scene normalization utilizing statistically invariant features, Proceedings of Workshop Atmospheric
Correction of Landsat Imagery, Defense Landsat Program Office,
[dates of workshop] Torrance, California, pp. 155-159.
Schott, J.R., C. Salvaggio, and W.J. Volchok, 1988. Radiometric scene
normalization using pseudoinvariant features, Remote Sensing of
Environment, 26:l-16.
35:11-27.
- - .- .
Henebry, G.M., and H. Su, 1993. Using landscape trajectories to assess Singh, A., 1989. Digital change detection techniques using remotelysensed data, International Journal of Remote Sensing, 10(6):
the effects of radiometric rectification, International Journal of
989-1003.
Remote Sensing, 14(12):2417-2423.
Teillet,
P.M., 1986. Image correction for radiometric effects in remote
Hill, J., and B. Sturm, 1991. Radiometric correction of multitemporal
sensing, International Journal of Remote Sensing, 7(12):
Thematic Mapper data for use in agricultural land-cover classifi1637-1651.
cation and vegetation monitoring, International Journal of Remote
Thome, K., B. Markham, J. Slater, and S. Biggar, 1997. Radiometric
Sensing, 12(7):1471-1491.
calibration of Landsat, Photogmmmetric Engineering 6 Remote
Jensen, J.R., 1983. Urban and suburban land use analysis, Manual of
Sensing, 63(7):853-858.
Remote Sensing, Second Edition, Vol. II, Interpretation and AppliYuan,
D., and C.D. Elvidge, 1996. Comparison of relative radiometric
cations (John R. Estes, editor), ASPRS, Falls Church, Virginia,
normalization techniques, ZSPRS Journal of Remote Sensing,
pp. 1618-1620.
51:117-126.
, 1996. Introductory Digital Image Processing, Prentice Hall,
[City], New Jersey, 316 p.
(Received 29 October 1998: revised and accepted 07 September 1999)
-
-
Certification Seals & Stamps
Now that you are certified as a remote
sensor, photogrammetrist or GIS/LIS
mapping scientist and you have that
certificate on the wall, make sure everyone
knows!
ASPRS Certification Seals 6 Stamps. 5410 GI-or
Lane. S u ~ t 210.
e Bethesda. MD 20814-2160
NAME
PHONE
certified finishing touch to your professional product.
CERTIFICATIONx
EXPIRATIONDATE
You can't carry around your certificate, but
your seal or stamp fits in your pocket or
briefcase.
ADDRESS
An embossing seal or rubber stamp adds a
W
W
I
980
SEND COMPLETED FORM WITH YOUR PAYMENT TO:
To place your order, fill out the necessary mailing and certification information.
Cost is just $35 for a stamp and $45 for a
seal; these prices include shipping and
handling. Please allow 3-4 weeks for
August 2000
CITY
STATE
PLEASE SEND ME:P Embossing Seal .......... $45
POSTAL CODE
I
COUNTRY
q Rubber Stamp $35
METHOD OF PAYMENT: 0 Check 0 Visa 0 Mastercard 0 American
CREDIT CARD ACCOUNT NUMBER
EXPIRES
SIGNATURE
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING