Image-Based Atmospheric Corrections -

Revisited and Improved

Pat S. Chavez, Jr.

Abstract as mapping land use, geology, forest types, etc. Multitem-

A major benefit of multitemporal, remotely sensed images i s poral image data collected, and being collected, by various
their applicability to change detection over time. Because of imaging systems are becoming increasingly important be-
concerns about global and environmental change, these data cause of concerns about global and environmental change. A
are becoming increasingly more important. However, to max- major benefit of multitemporal remotely sensed image data is
imize the usefulness of the data from a multitemporal point its applicability to change detection (Robinove et al., 1981;
of view, an easy-to-use, cost-effective, and accurate radiomet- Jensen and Toll, 1982; Fung, 1990; Chavez and MacKinnon,
ric calibration and correction procedure is needed. The at- 1994). However, to maximize the benefit, change detection
mosphere effects the radiance received at the satellite by must be done automatically in an operational environment,
scattering, absorbing, and refracting light; corrections for and the results must be related to physical ground units.
these effects, as well as for sensor gains and offsets, solar ir- This implies an operational radiometric-correction procedure
radiance, and solar zenith angles, must be included in radio- for both sensor and atmospheric effects. There are several
metric correction procedures that are used to convert different methods to convert the image digital counts (DCS) to
satellite-recorded digital counts to ground reflectances. To reflectances. The most straightforward method converts Dcs
generate acceptable radiometric correction results, a model is to so-called apparent or at-satellite reflectances by correcting
required that typically uses in-situ atmospheric measure- for sensor gains, offsets, solar irradiance, and solar zenith an-
ments and radiative transfer code (RTC)to correct for atmo- gle (Markham and Barker, 1986; Price, 1987; Leprieur et al.,
spheric effects. The main disadvantage of this type of 1988; Hall et a]., 1988; Hall et al., 1989; Chavez, 1989;
correction procedure is that it requires in-situ field measure- Moran et al., 1992). However, that method does not correct
ments during each satellite overflight. This is unacceptable for atmospheric effects, which can be substantial.
for m a n y applications and i s often impossible, as when using The amount of electromagnetic energy sensed by an im-
historical data or when working in very remote locations. aging system's detectors is influenced by the atmosphere. At-
The optimum radiometric correction procedure i s one mospheric effects are wavelength dependent; are both
based solely on the digital image and requiring n o in-situ additive and multiplicative in nature; and include scattering,
field measurements during the satellite overflight. The dark- absorption, and refraction of light (Curcio, 1961; Turner et
object subtraction (DOS) method, a strictly image-based tech- al., 1971; Sabins, 1978; Slater et al., 1983). Various methods
nique, is an attempt to achieve this ideal procedure. to remove the additive scattering component caused by path
However, the accuracy i s not acceptable for m a n y applica- radiance have been developed, including simple image-based
tions, mostly because it corrects only for the additive scatter- dark-object subtraction (DOS) techniques (Vincent, 1972;
ing effect and not for the multiplicative transmittance effect. Chavez, 1975; Ahern et a]., 1977; Chavez, 1988; Chavez,
This paper presents an entirely image-based procedure that 1989). However, atmospheric transmittance, which has a
expands on the ~10s model b y including a simple multiplica- multiplicative effect caused by both scattering and absorp-
tive correction for the effect of atmospheric transmittance. tion, does not have an equivalently simple and straightfor-
Two straightforward methods to derive the multiplicative ward correction procedure. To accurately correct for the
transmittance-correction coefficient are presented. The multiplicative effect due to transmittance usually requires in-
COSITZ) or COST method uses the cosine of the solar zenith situ field measurements of atmospheric optical depth (Slater,
angle, which, to a first order, i s a good approximation of the 1985; Slater, 1988; Holm et al., 1989). A simple but accurate
atmospheric transmittance for the dates und sites used in atmospheric correction procedure is needed, not only for op-
this study. The default TAUS method uses the average of the erational applications but also for many research projects
transmittance values computed by using in-situ atmospheric that use historical data sets. The full potential of historical
field measurements made during seven different satellite over- and new data will not be realized until an accurate, easy-to-
flights. Published and unpublished data made available for use, and cost-effective radiometric-correction procedure is
this study b y Moran et al. (1992) are used, and m y model re- developed that converts image Dcs to ground reflectances.
sults are compared with their results. The corrections gener- The optimum atmospheric correction procedure would
ated b y the entirely image-based COST model are as accurate be based solely on the digital image and would require no
as those generated b y the models that used in-situ atmo- in-situ field measurements during the satellite overflight. The
spheric field measurements and RTC software. DOS method, a strictly image-based technique, is an attempt
toward this ideal procedure; however, the accuracy is not ac-
Introduction
Multitemporal images collected by digital multispectral imag- Photogrammetric Engineering & Remote Sensing,
ing systems have been available for more than 20 years and Vol. 62, No. 9, September 1996, pp. 1025-1036.
have been used for various Earth-science applications, such
0099-1112/96/6209-1025$3.00/0
U.S. Geological Survey, 2255 N. Gemini Drive, Flagstaff, AZ O 1996 American Society for Photogrammetry
86001. and Remote Sensing

PE&RS September 1996 1025

 TABLE1. DATES,DISTANCES, AND TOP OF THE ATMOSPHERE on 5 s (Tanre et al., 1990). These models corrected for both
SOLARSPECTRAL
IRRADIANCES.DOY, DAYOF YEAR;D, EARTH-SUN DISTANCE IN ASTRONOMICAL
the additive scattering and the multiplicative transmittance
UNITS;Eo1, 2, 3, 4 ARE LANDSAT TM BANDS1,2, 3, AND 4 TOPOF THEcomponents, while the 110smodel corrected only for the ad-
ATMOSPHERE SOLARSPECTRAL IRRADIANCE(W M - ~V M - ~ ) MODIFIED
, TO INCLUDE
ditive scattering component due to path radiance. This paper
THE AFFECTOF THE D * D TERMIN EQUATION 2 (MORAN,PERSONAL
presents a comparison among the results generated by (1) the
COMMUNICATION), FOR SEVENSATELLITE OVERFLIGHT DATES(MORANETAL.,
1992). in-situ atmospheric optical depth measurements RTC method
(HBC, the best), (2) RTC with simulated atmosphere method
Date DOY D Eol Eo2 Eo3 Eo4 ( ~ s D )(3)
, the DOS method, and (4) the new COST and default
23 Jul 85 204 1.01580 1895.1 1770.5 1497.3 1010.6 TAUS entirely image-based models derived in this paper.
08 Aug 85 220 1.01382 1895.1 1770.5 1497.3 1010.6 Moran et al. (1992) used the University of Arizona's
27 Oct 85 300 0.99360 1980.8 1850.5 1565.0 1056.3 Maricopa Agricultural Center (MAC) as their test site over
20 Mar 86 079 0.99608 1970.9 1841.3 1557.2 1051.0 which the simultaneous ground, aircraft, and satellite mea-
05 Apr 86 095 1.00065 1953.0 1824.5 1543.0 1041.4 surements were made (Jackson, 1990). Landsat-5 Thematic
21 Apr 86 111 1.00515 1935.5 1808.2 1529.2 1032.1 Mapper (TM) data collected during overflights from April
24 Jun 86 175 1.01690 1891.0 1766.7 1494.1 1008.4 1985 to June 1986 were used (Moran, 1986). Twelve TM
scenes were collected, but five of them were not used be-
cause of unfavorable cloudy conditions. The satellite over-
ceptable for many applications, especially those dealing with pass specifications for the seven dates used are shown in
medium to bright reflectance values. One of the main reasons Table 1. The field and aircraft measurements include seven
for the unacceptable accuracy is because the ~ 0 model
s does soil and seven vegetation ground targets, which are shown as
not correct for the multiplicative effect of transmittance. This separate groups in the tables and bar graphs. However, I
paper's main objective is to present an entirely image-based moved the vegetation entry for 2 7 October 1985 (DOY 300)
procedure that expands on the DOS model by including a into the soils group because of its moderately higher reflec-
simple multiplicative correction for transmittance. Published tance values in the visible bands and lower vegetation index.
and unpublished data from Moran et al. (1992) were made These two catogories are used to generate various statistical
available to the author for this study. Therefore, this paper information for comparisons throughout the paper. Table 2
also compares their results with results generated by the new shows results of the aircraft-based reflectance measurements
COST and default TAUS image-based models. for the various soils and vegetation targets. On the basis of
earlier work. Moran et al. 119921 assumed that these values
Background and Data Set Used were equivalent to ground reflectances and used them to de-
As stated by Moran et al. (1992), "in response to the need for termine the accuracy of the various models (Holm et al.,
a simple atmospheric correction method and the consequent 1989). To make it easier to compare results, several tables
verification of such a method, an experiment was designed and plots in this paper duplicate the information given by
to acquire a data set suitable for testing atmospheric correc- Moran et al. (1992). For a detailed description of the proce-
tion procedures under a variety of conditions." With Dr. dures used for ground, aircraft, and atmospheric measure-
Moran's ~ermission,I have used the same data set in my ments and complete analysis, refer to Moran et al. (1992).
analYsis.&shemade available not only the data published in
the paper, but also other data used to derive those shown in Radiometric Correction Models
their tables and figures. This made the evaluation and corn- Generally, the objective of a radiometric atmospheric correc-
parison of the new image-based models easier and more tion procedure is to convert satellite-generated digital counts
complete. (DCS)to ground reflectances (i.e., absolute surface reflec-
To evaluate the results generated by the various models, tance~).How the different model parameters are derived de-
Moran et al. used spectral data collected simultaneously by pends on the available information (i.e., ground and/or
satellite, aircraft, and ground-based sensors over large uni-
form ground targets during multiple satellite overflights. At-
mospheric optical depth measurements were also made TABLE2. AIRCRAFT-BASED MEASURED REFLECTANCES OF MORANETAL. (1992).
during each satellite overflight and used as input to radiative ON THE BASISOF PREVIOUS WORK,REFLECTANCES AREASSUMEDTO REPRESENT
transfer codes (RTCS)to compute surface reflectances. As ex- THE GROUNDREFLECTANCES AND ARE USEDTO CHECKTHE ACCURACY OF THE
petted, the results derived by using the in-situ optical depth VARIOUSMODELS.GROUND TARGETSARESEPARATED INTO SOILSOR VEGETATION

measurements and RTC software were the most accurate. AS THEYWEREBY MORANETAL. DOY, DAYOF YEAR,1985-86.

They compared these results with those generated by using (1985-86) TM1 TM2 TM3 TM4
simpler atmospheric correction models. Their evaluation in-
cluded use of the in-situ atmospheric measurements as input Soils
to different RTCs: Herman-Browning (HBC)(Herman and 204 0.0805 0.1205 0.1684 0.2155
Browning (1965), which gave the best results) and 5s (Tanre 220 0.0845 0.1256 0.1854 0.2250
et al., 1990). However, as stated in their paper, "This type of 300 0.0745 0.0941 0.1082 0.2425
procedure has been shown to be accurate by Holm et al. 300" 0.0953 0.1424 0.1947 0.2437
(1989) and Moran et al. (1990), but to expensive and time- 079 0.0698 0.1072 0.1539 0.1997
consuming for it to be considered for use operationally." 095 0,0664 0.0992 0.1498 0.2039
The simpler image-based methods used by them for 111 0.0924 0.1398 0.2061 0.2736
17' 0.1046 0.1491 0.2101 0.2662
comparison included using (1)RTCs with simulated, rather
Vegetation
than measured, atmospheric information and (2) the dark-ob-
ject subtraction (DOS)technique (Vincent, 1972; Chavez, 204 0.0232 0.0589 0.0269 0.5387
1975; Chavez, 1988; Chavez, 1989). The model that used a 220 0.0302 0.0591 0.0365 0.5861
simulated atmosphere to infer the required correction para- 079 0.0260 0.0465 0.0302 0.3758
meters was suggested by Ahern et 01. (1977). The informa- Og5 0.0370 0.0634 0.0533 0.4124
0.0235 0.0542 0.0239 0.6258
tion derived by using the simulated atmosphere was used
with a simple procedure based on LOWTRAN 7 (Kneizys et
ii: 0.0310 0.0650 0.0336 0.5713
al., 19881, as well as with a more complex procedure based *Moved f r o m vegetation section.

September 1996 PE&RS I

atmospheric in-situ measurements or lack thereof). Regard- The main advantange of this model is that it is very simple
less of the model, though, DCS must first be converted to at- and easy to apply; it does not require in-situ field measure-
satellite radiances by removing the gain and offset effects ments of ground reflectances andlor atmospheric optical
introduced by the imaging system. If the data have been depths or simulated atmospheric parameters. The disadvan-
processed to remove striping noise by using a statistical tech- tages are that the accuracy is usually not acceptable for many
nique, these additional gains and offsets must also be in- applications, especially for TM bands 1 and 4 (Moran et a].,
cluded in the correction. The gain and offset values used for 1992), and the correction is based on the sensor's gain and
these particular data are given in Moran et al. (1992, p. 175) offset values, which are uncertain due to changing sensor
and are not duplicated here. The equation to convert satellite characteristics (Slater et al., 1987; Thome et al., 1993; Moran
DCS to at-satellite radiances is et al., in press). This second disadvantage is one that exists
for all the models considered in this paper and should be
Lsat = (DC - 0ffset)lGain (1) kept in mind by the reader.
where Lsat is the at-satellite spectral radiance for the given
spectral band (W m-2 sr-1 ym-l), DC is the digital count at DOS Model
the given pixel for the given spectral band, Offset is the off- The image-based dark-object subtraction (DOS) model used in
set for the given spectral band (DC), and Gain is the gain for the comparison by Moran et al. (1992) is the one discussed
the given spectral band (DC mZsr ym W-I). in several previous papers (Vincent, 1972; Chavez, 1975;
Chavez, 1988; Chavez, 1989). The basic assumption is that
Next, at-satellite radiances must be converted to surface within the image some pixels are in complete shadow and
reflectances by correcting for both solar and atmospheric ef- their radiances received at the satellite are due to atmo-
fects. The general modellequation used to do this and pre- spheric scattering (path radiance). This assumption is com-
sented as Equation 3 by Moran et al. (1992, p. 172) is bined with the fact that very few targets on the Earth's
surface are absolute black, so an assumed one-percent mini-
(PI *
(Lsat - Lhaze)) mum reflectance is better than zero percent. The paper by
REF = (2)
(TAUV * (Eo * Cos (TZ) * TAUz + Edown)) Chavez (1989) discusses an improved method of selecting the
dark-object haze values for the separate spectral bands. The
The definitions are those given by Moran et al. (1992): i.e.. objective of that improved dark-object method is to select
spectral-band haze values that are correlated to each other,
REF Spectral reflectance of the surface.
= rather than by using the histograms of each spectral band in-
Lhaze Upwelling atmospheric spectral radiance scat-
= dependently, which can cause haze-selection problems when
tered in the direction of and at the sensor en- topographiclshadow conditions are minimal. Both methods
trance pupil and within the sensor's field of view generate haze values that are very similar when sufficient to-
(W m-2 sr-I km-l), i.e., the path radiance. pography exists, as in the images used for this study. The
TAUv = Atmospheric transmittance along the path from histogram and DOS one-percent haze-selection method was
the ground surface to the sensor. used by Moran et al. (1992); therefore, it is the one compared
Eo = Solar spectral irradiance on a surface perpendicu- in this paper.
lar to the sun's rays outside the atmosphere (W Besides correcting for the same parameters that the ap-
m-Z ym-I). Eo contains the Earth-sun distance parent reflectance model does, the image-based DOS radio-
term (D*D)imbedded and is in astronomical metric correction model also corrects for the atmospheric
units (AUS are a function of time of year and additive scattering component attributed to the path radi-
range from about 0.983 to 1.017). ance. Therefore, in the general radiance-to-reflectance model
TZ = Angle of incidence of the direct solar flux onto shown in Equation 2, the following applies for the Dos
the Earth's surface (solar zenith angle, Thetaz). model:
TAUz = Atmospheric transmittance along the path from
the sun to the ground surface. TAUz = 1.0 (ignores atmospheric transmittance),
Edown = Downwelling spectral irradiance at the surface TAUv = 1.0 (ignores atmospheric transmittance),
due to scattered solar flux in the atmosphere (W Edown = 0.0 (ignores downwelling), and
m-z ym-I). Lhaze = value derived from the digital image using the
dark-object criteria.
All the radiometric correction procedures below start
with this general model (Equation 2) but make different sim- The DOS model's main advantages are that it is strictly an
plifying assumptions that eliminate certain parameters. The image-based procedure and does not require in-situ field
information available about the data and atmospheric condi- measurements and that it is simple and relatively straightfor-
tions determines what assumptions must be made and, there- ward to apply. The main disadvantages are that for reflec-
fore, the specific model that is to be used. Often, the method tance values greater than about 15 percent the accuracy is
used to derive the required parameters can also determine often not acceptable and that the selection of the haze values
assumptions that must be made. must be done with care.

Moran eta/.Models
Apparent Reflectance Model
The apparent reflectance model is the simplest one used to The evaluation by Moran et al. (1992) examined several dif-
convert at-satellite radiances to reflectances. It corrects for ferent models to derive the atmospheric parameters needed
for Equation 2. These models included using RTC solutions
spectral band solar irradiance and solar zenith angle of the
generated by using both in-situ measured (HBC and 5s) or
image but makes no attempt to correct for atmospheric scat-
tering and absorption. For the apparent reflectance model, simulated (LOWTRAN 7 and 5s) atmospheric information. The
comparisons in this paper are limited to the uncorrectedlap-
the following applies in Equation 2:
parent, DOS 1-percent, HBC (which gave the best results), and
TAUz = 1.0 (ignores atmospheric transmittance), simulated 5s models, and the new entirely image-based mod-
TAUv = 1.0 (ignores atmospheric transmittance), els derived in this paper. As expected, the most accurate
Edown = 0.0 (ignores downwelling), and model in the Moran et al. (1992) study was the one that used
Lhaze = 0.0 (ignores scattering due to path radiance). in-situ atmospheric measurements, made during the satellite

PE&RS September 1996

TABLE3. SOLAR ZENITHANGLES, THEIR COSINES,OEL VALUES,AND COMPUTED TAUZ VALUES FOR TM BANDS1,2, 3, AND 4 FOR THE 1985-86OVERFL~GHT DATES.
DOY, DAYOF YEAR FOR SATELLITE
OVERFLIGHT-IMAGEDATE;THETAZ (TZ), SOLAR
ZENITHANGLE; COS(TZ),COSINEOF THETAZ; DEL, OPTICALDEPTH VALUES COLLECTED
TAUZ VALUES; AVE, AVERAGETAUZ VALUES.
BY MORANETAL.; TAUZ, MULTIPLICATIVE TAU2 VALUES WERE COMPUTED USING THE OPTICALDEPTH MEASUREMENTS
GIVENBY MORANETAL. FUNCTION USED TO COMPUTETAUZ WAS EQUALTO EXP(-OEL * SEC(THETAZ)), WHERE DEL IS EQUALTO OPTICALDEPTH, EQUATION3.
Ave ,
DOY THETAZ COS(TZ) dellITAUz1 delZITAUz2 del31TAUz3 del4lTAUz4 TAUz
204 29.84 0.87 0.29/0.71 0.22/0.78 0.1410.85 0.07/0.92 0.82
220 30.08 0.87 0.3310.68 0.22/0.78 0.1610.83 0.1010.89 0.80
300 50.88 0.63 0.2910.64 0.20/0.73 0.1310.81 0.0710.89 0.77
079 43.76 0.72 0.2510.71 0.1810.78 0.11/0.86 0.0510.93 0.82
095 38.35 0.78 0.25/0.72 0.20/0.77 0.12/0.86 0.06/0.93 0.82
111
--- 33.35 0.84 0.21/0.78 0.1310.86 0.09/0.90 0.05/0.94 0.87
175 27.56
Ave.
Note: Table 3 shows that, to a first order, cosine(thetaz)is equal to the TAUz value of the individual spectral bands. Average values were
computed to show possible default values for each spectral band; these are the values used by the default TAUS model. The del values are
rounded off to two places, but the actual computations were done using four place accuracy.

overflights, which included optical depths as input to RTC eral radiance-to-reflectance model (Equation 2) to compute
software to compute the relation between surface reflectance surface reflectances.
and sensor radiance (Moran et al., 1992, p. 170). However, The first method used to derive TAU values independent
acceptable results were also generated with the model that of in-situ field measurements is very straightforward. As
used simulated atmosphere parameters as input to RTC soft- given by Moran et al. (1992, p. 172), the multiplicative trans-
ware. A main difference between the models using RTC soft- mittance component for scattering and weak absorption is
ware and the DOS method was that in-situ measurements or approximated by Equations 3 and 4 (repeated below):
simulated atmospheres were used to derive the required at-
mospheric parameters for Equation 2, including corrections TAUz = ~ ~ ~ ( - *d sec(Tz))
e l (3)
for transmittance and downwelling. In the general radiance- and
to-reflectance model shown in Equation 2, the following cri-
teria apply: TAUv = ~ ~ ~ ( - *d sec(Tv))
e l
TAUz = ~ ~ ~ ( - *d sec(TZ))
e l where del is the optical thickness values at given wave-
lengths, TZ is the solar zenith angle (thetaz), and TV is the
and viewing angle (zero degrees for Landsat TM images - thetav).
TAUv = ~ x ~ ( - d e* lsec(TV)) Using the spectral optical depth and solar zenith values
given by Moran et al. (1992, p. 174), the TAUz values were
where EXP is the exponential, sec is the secant, del is the op- computed for TM bands 1, 2, 3, and 4. Table 3 shows the
tical thickness values measured in-situ at the given wave- computed TAUZ values, along with solar zenith angles and
lengths, TZ is the solar zenith angle, thetaz, and TV is the their cosines for the seven satellite overflight dates. Empiri-
viewing angle (zero degrees for nadir viewing systems, thetav). cally, the values in Table 3 show that, to a first order, the co-
Edown is derived from atmospheric measurements made in- sine of the solar zenith angle is a good approximation of
situ or computed from simulated atmospheres, and Lhaze is TAUz. It approximates TM bands 1 and 2 best in some cases
derived from atmospheric measurements made in-situ, or and TM bands 3 and 4 best in others. However, the average
simulated atmosphere, or by using the DOS Lhaze procedure. TAUz of all bands for all dates is very similar to the average
The advantage of the RTC-drivenmodels was the im- of the cosine values (0.81 and 0.80, respectively). Equation 3
proved accuracy. The procedure that used simulated atmo- shows that transmittance is a function of the solar zenith an-
spheres rather than in-situ measurements had the added gle (TZ) and the optical depth (del). For most acceptable im-
advantage of not requiring someone to be in the field during ages, TZ is in the range of 30 to 55 degrees and del has a
the satellite overflight. However, it did require use of the RTC range of 0.08 to 0.30. Therefore, in the EXP(-del * sec(T2))
software. The HBC model generated the best results but re- function, the variation of sec(TZ) is about 2.7 times larger
quired in-situ measurements and, therefore, someone in the than that for del, which implies that Tz (the solar zenith an-
field during each of the satellite overflights. gle) is the more dominant variable. So, a relation that is
strictly dependent on TZ to approximate the exponential
Improved Image-Based Model function, to a first order, may be acceptable (i.e., set it equal
The improvement made to the image-based DOS model is to the cosine of TZ). This empirically observed relation was
based on a method that derives a correction for the multipli- used as the first method to select TAUZ values.
cative transmittance effect by using one of two techniques. To help strengthen this empirically observed relation
Equation 2 shows that the error for not including a multipli- seen in Table 3 between the cosine of the solar zenith angle
cative correction for transmittance is approximately 1.0/ (TZ) and EXP(-del*sec(~z)), the power series expansion of
(TAUZ * TAUV) which, for the seven dates used, implies an these two functions was used. The first four terms of the
approximate overall error of 30 percent. Therefore, a correc- power series are
tion for the multiplicative transmittance component can sub-
stantially improve the DOS model results. In this study two
different methods were used to derive the required TAU val-
ues in Equation 2. Both methods that correct for multiplica- and
tive transmittance effects are independent of in-situ
atmospheric and ground measurements. Field-independent
derived TAU values were used, along with the DOS Lhaze ad-
ditive-scattering component due to path radiance, in the gen- The solar zenith angle TZ is in radians and ! represents fac-

September 1996 PE&RS

toral; the power series expansions were done using the stan- corrected for dark reflectances and under corrected for bright
dard Taylor series equations (Boas, 1966, p. 19). The average reflectances, while the DOS model tends to under correct
TZ and del values for the seven dates used are 0.64 and 0.17, most of the reflectances (see Figure 1).Therefore, the appar-
respectively. Using these two average values in the power se- ent reflectance model has a larger number of negative and
ries expansions generates values of 0.8021 for the cosine and positive errors that cancel each other out compared to the
0.8087 for the exponential when carried out four terms; if DOS model. Because the question of over or under correction
only the first two terms are used, the cosine value is 0.7952 is important, the tables show differences between the com-
and the exponential value is 0.7882 (both are within one per- puted and measured values for each band on every date in
cent of each other). As in Table 3, these calculations also positive and negative format. The graphs show the scatter
shows that, to a first order, the cosine of the solar zenith an- plots of all the measured versus computed values for each
gle and the exponential of minus del times the secant of the model and show whether the data are being over or under
solar zenith angle are equal for these data. corrected (see Figures 1 and 2).
As will be shown in the results section, using the cosine Table 2 shows the aircraft-based measured-reflectance
of the solar zenith angle for TAUz substantially improves the values that Moran et al. (1992, p. 174) considered to be the
DOS results, and, in fact, the results are as good as those gen- correct ground readings. As in Moran et al. (1992), this paper
erated by the HBC model. I call the cosine of the solar zenith maintains the same groups of soils and vegetation except, as
angle, COS(TZ), correction the COST model. noted earlier, the vegetation entry for DOY 300 was moved
Because the cosine is independent of wavelength, an al- into the soils group because both the moderately high reflec-
ternative default method, one dependent on wavelength, was tance values in the visible bands and the lower vegetation
also developed for comparison. It is also quite simple, index imply a soils target. In TM bands 1, 2, and 3 (visible
straightforward, and entirely image based. The alternative bands), the reflectances of these two groups also correspond
method simply uses default TAUz values, which are the aver- closely to dark versus bright reflectances. Besides the com-
age for each spectral band using the seven dates given by the puted reflectance values for the given model, each table
Moran et al. (1992) data. These default values for TM bands shows the difference between the computed and the meas-
1, 2, 3, and 4 are the averages given in Table 3. I refer to this ured (Table 2) values. In these tables, positive implies over
correction as the default (DEF) TAUS model and can be used corrected and negative implies under corrected results for
instead of the COST model when information on a per image the following models:
basis is not available. In the general radiance-to-reflectance
model shown in Equation 2, the following criteria apply: Table 4, Uncorrectedlapparent reflectance
Table 5, DOS 1-percent reflectance
TAUz = Cos(thetaz) for the COST model Table 6, 5SD reflectance
or Table 7, HBC reflectance
TAUz Default TAUz values from Table 3 for the DEF
= Table 8, COST reflectance
TAUS model; 0.70, 0.78, 0.85, and 0.91 for TM Table 9, Default TAUz reflectance
bands 1, 2, 3, and 4, respectively;
TAUv = Cos(thetav) = 1.0 because thetav is zero degrees Tables 4, 5, 6, and 7 give the results generated by Moran
for nadir view; et al. (1992); Tables 8 and 9 show the results generated by
Edown = 0.0 (ignores downwelling); and the new COST and default TAUS image-based models. The ta-
Lhaze = Value derived from the digital image using the bles can be used to make very detailed comparisons between
dark-object criteria (identical to the DOS model). the various models. Scatter plots like those used by Moran et
al. (1992) are also shown for general comparisons (Figures 1
As in the DOS model, the main advantages of these mod- and 2). The plots in Figure 1 are identical to those published
els are that they are strictly image-based procedures and by Moran et a1. (1992). The HBC model used the in-situ at-
therefore do not require in-situ field measurements and they mospheric measurements as input to RTC software and, as
are simple and very straightforward to apply. Compared to expected, generated the best results. The 5SD model, which
that of the DOS model, the accuracy generated by both the used simulated atmospheric information as input to RTC soft-
COST and DEF TAUS image-based models is substantially im- ware, also generated acceptable results, but the DOS 1-percent
proved, and use of RTC software is not required. However, as model was good only for the darker reflectances. The scatter
for the Dos model, a disadvantage is that the dark object DN plots in Figure 2 show the results of the improved image-
value for Lhaze must be selected carefully. based models that derive the multiplicative transmittance
correction coefficient by using either the cosine of the solar
Results and Discussions zenith angle (COST) or the default TAUZ values (DEF TAU^).
This section compares the aircraft-based measured reflec- From the scatter plots and from Tables 8 and 9, we can see
tance~,which are used by Moran et al. (1992) as equivalent that the two new models have substantially improved on the
ground reflectances based on previous research (Holm et al., DOS results; in fact, the COST model generated results as good
1989), with the computed values generated by the various as the HBC (best) model. The default TAUZ model also im-
models. As done by Moran et al., the results are compared in proved the DOS results, but not as much as the COST model.
table, plot, and bar-graph format. The tables show the differ- The two sets of bar graphs in Figure 3 compare the re-
ences between the computed reflectance values and the air- flectance accuracy of each model for each spectral band for
craft-based ground-measured values, as well as the average of the soils group and for the vegetation group. Generally, this
the absolute value of these differences for each spectral band is a bright and dark reflectance comparison except for the
on each satellite overflight date. Note that these comparisons near-infrared TM band 4 where vegetation reflectance values
differ slightly from those made in several graphs presented are high. For each model, the graphs show the absolute dif-
by Moran et al. (1992). Several of their graphs show the aver- ference between the computed and measured reflectances for
age of the difference, not of the absolute value of the differ- each individual band as well as the average of all the TM
ence. The absolute-value method emphasizes the amplitude bands. This allows a comparison of not only the overall ac-
of the error rather than whether the overall results are over curacy of each of the models, but also the accuracy within
or under corrected, and it keeps positive and negative errors individual TM bands for the two data groups. For soils,
from canceling each other out. For example, the uncorrected1 which have mostly mid to high reflectance values, both the
apparent reflectance model generates results that are over COST and default TAUz models improve on all the DOS re-

PE&RS September 1996

 @!
2
E
0 2?
t= m
S
E
m
m\
u U
K
a, 0
2
> 2?
2L &O
P 5
B a,
+j5
X
m C Q
;5m
Y
.=0 u
-- 2a, 2
-
-7
8
V)
a
u
P -a PE
U M
a,

8 2 z4E 63
r --m
3
V)
'a
;2
a,
55
WE
.G a,
m V)
3 P
b Q

o
C -0
m c
m
c-6,
- (U
g
F X 2 X 2 ;
0
dd
m
U
-
i
; . I . !
0 0 X Z X Z " 0 X 0

8c -,
(0
t l soa
e 3 u a ~ a l ~ e(%L) a3ue13elbatl ~ E I H
F
g 20
?
2
0
32
C K

(D .
. ?
E czl
a,
c m
P
0

1 a,

. ;
"I
m
Q

fc -
5
2
2 -- z FE
.-2 'Z
-D
ad
uU
C
a,

: m % 2? a, uj

-u2 2 a * = ;2 5
2 =Qm \
a
-- 6
-Q 8 $:
3
g0
=
u
g
a
,
m
m
C
rn
$ 2
a
- E E. uE
z 2-
(JY

c 2 2E .
. ? 2E @$
a 0
r
.-ae 'a u m o
$ t E
m a
LlmX
Y
0
i:, .
. Y bas
.s
-d a
ag2
:42
oe-
E z z g S"rE
....I
7

o
28 8
E
om t
m a
0 6,
dd
I I I 4
0
F G ~
L X z : z z ; O 2 ? z s
0 o0 "0 -0 o Gm
aouetsa~ianpasaq-at!lletes a a u e ~ o a ~ ~ass
atl iT (U-

1030 September 1996 PE&RS

 COST Default TAUS

0 7

TM1
0 6 --
TMZ

0 5 TM3
C
I
+
w TM4
%
0:
04--

E,-
O 0 3
$
$

3 0 1 0 2 0 3 04 05 06 I 0 0 1 0 2 0 3 0 4 0 5 0 6 0 7

Aircraft-based Reflectance A~rcraft-basedReflectance

(a) (b)
F~gure2. Results generated by the COST (a) and default TAUz (b) models on computed satellrte-based versus measured aircraft-based reflectances for TM bands 1 , 2,
3, and 4 uslng the soils and vegetation data for seven dates from 23 July 1985 to 2 4 June 1986 (Table 1).

SOILS VEGETATION

UNC DOS 5SD HBC COST DEF TAUS

UNC DOS 5SD HBC COST DEF TAUS

(a) (b)
Figure 3. Reflectance results for spectral-bands T M 1,2, 3, and 4 and an overall average, as determined by six models for the soils (a) and vegetation (b) targets. The
average of the absolute difference between the computed and aircraft-based measured values are plotted.
 TABLE4. COMPUTED RERECTANCES GENERATED BY USINGTHE SATELLITE cause the reflectance of vegetation in an agricultural
UNCORRECTED/APPARENT REFLECTANCE MODEL(MORANETAL., 1992), THE environment can be much higher than the background soils,
DIFFERENCESBETWEEN THE COMPUTED AND AIRCRAFT-BASED GROUNDMEASURED so an error in ground sampling would be more pronounced.
VALUES,AND THE AVERAGE OF THE ABSOLUTE DIFFERENCE BETWEENTHESE In the visible bands the error would not be as noticable be-
VALUES.DOY, DAYOF YEAR;AVE, AVERAGE; DIFF, DIFFERENCE.
cause the reflectance level of vegetation is more similar to
DOY TM1 TM2 TM3 TM4 Ave/Diff/ the background soils. In a previous paper it was shown that
in densely vegetated areas the near-infrared band generally
Soils
0.1603 0.2098
has more spatial variability than do the visible bands
204 0.1293 0.1307
Diff 0.0488 0.0102 -0.0081 -0.0057 0.0182 (Chavez, 1992); therefore, it may be more difficult to accu-
220 0.1281 0.1255 0.1539 0.1986 rately sample vegetated targets.
Diff 0.0436 -0.0001 -0.0315 -0.0264 0.0254 One thing that stands out from the vegetation bar graphs
300 0.1370 0.1176 0.1195 0.2175 is the fact that the DO^ model does better than any of the
Diff 0.0625 0.0235 0.0113 -0.0250 0.0306 others, including both the HBC and the COST models, for the
300* 0.1449 0.1457 0.1804 0.2253 visible bands (TM bands 1 , 2, and 3). However, it does poorly
Diff 0.0496 0.0033 -0.0143 -0.0184 0.0214 on TM band 4, as stated by Moran et al. (1992), which can be
079 0.1189 0.1199 0.1481 0.1956 due to several reasons, including (1)the lack of a transmit-
Diff 0.0491 0.0127 -0.0058 -0.0041 0.0179
095 0.1147 0.1110 0.1367 0.1811
tance correction in the model, (2) possible aircraft-based re-
Diff 0.0483 0.0118 -0.0131 -0.0228 0.0240 flectance measurement errors, and (3) the fact that TM band-4
111 0.1319 0.1411 0.1829 0.2380 reflectance values for vegetation are quite high, so we are
Diff 0.0395 0.0013 -0.0232 -0.0356 0.0249 seeing a bright, not dark, target error condition. The similar
175 0.1467 0.1564 0.1971 0.2525 shapes of the bar graphs for 5SD, HBC, and COST imply that
Diff 0.0421 0.0073 -0.0130 -0.0137 0.0190 errors remaining in these three models are similar be., they
Ave/Diff/ 0.0479 0.0088 0.0150 0.0189 0.0227 are all correcting for the same errors, and errors that are left
Vegetation are common to all three models).
204 0.0940 0.0851 0.0548 0.4888 Several additional points should be made on the basis of
Diff 0.0708 0.0262 0.0279 -0.0499 0.0437 the data and the results seen in this study.
220 0.0999 0.0848 0.0615 0.4811 (1) The cosine-derived TAUZ values are a function o f time o f
Diff 0.0697 0.0257 0.0250 -0.1050 0.0564 year but are independent o f wavelength, w h i l e the default
079 0.0919 0.0734 0.0523 0.3618 TAUZ values are a function o f wavelength but independent
Diff 0.0659 0.0269 0.0221 -0.0140 0.0322 o f t i m e o f year. T h e TAUZ values in Table 3 show larger dif-
095 0.0999 0.0896 0.0742 0.3522 ferences in functions o f wavelength t h a n o f t i m e of year;
Diff 0.0629 0.0262 0.0209 -0.0602 0.0426
111 0.0925 0.0827 0.0547 0.5418
D iff 0.0690 0.0285 0.0308 -0.0840 0.0531 TABLE5. COMPUTED REFLECTANCESGENERATED BY USINGTHE IMAGE-BASED
175 0.1013 0.0937 0.0648 0.5179 DOS 1-PERCENT MODEL(MORANETAL., 1992), THE DIFFERENCES BETWEEN THE
Diff 0.0703 0.0287 0.0312 -0.0534 0.0459 COMPUTED AND AIRCRAFT-BASED
GROUNDMEASURED VALUES,AND THE AVERAGE
Ave/Diff/ 0.0682 0.0270 0.0263 0.0611 0.0457 OF THE ABSOLUTEDIFFERENCE
BETWEEN THESEVALUES.DOY, DAYOF YEAR,
* M o v e d f r o m vegetation section. 1985-86; AVE, AVERAGE;
DIFF, DIFFERENCE.
DOY(1985-86) TM1 TM2 TM3 TM4 Ave/Diff/

sults. In fact, the COST model does as well as or slightly bet- 204
ter than the HBC model. The default TAUz model still has a D iff
problem with the near-infrared TM band 4 but does improve 220
Diff
on the DOS model; the visible bands have acceptable accu- 300
racy, and results are similar to those generated by the COST Diff
and HBC models. In general, the D o s model has a problem 300*
with the bright soils, and results are better than the uncor- D iff
rected results for only TM band 1 (as stated by Moran et al. 079
(1992)). As shown later, this is not the case for the darklveg- Diff
etation targets. Notice that the shape of the bar graphs for 095
models 5SD, HBC, and COST are similar. Diff
111
The vegetation reflectance bar graphs, which for the visi-
Diff
ble bands represents mostly very low reflectance values, also 175
show that the COST model generates results similar to those Diff
of the HBC model; in fact, the overall average is better be- Ave/Diff/
cause of improved results for TM bands 2 and 4. The default
TAUz model generates acceptable results for the visible Vegetation
204 0.0285 0.0497 0.0329 0.4880
bands, but, like all the other models, the near-infrared band
Diff 0.0053 -0.0092 0.0060 -0.0507 0.0178
has larger errors. It is not clear why the near-infraredl~~ 220 0.0380 0.0538 0.0416 0.4865
band 4 consistently has larger errors in all the models for the Diff 0.0078 -0.0053 0.0051 -0.0996 0.0295
vegetation group. Part of the reason may be that the multipli- 079 0.0303 0.0403 0.0328 0.3170
cative correction does not fully account for the absorption ef- D iff 0.0043 -0.0062 0.0026 -0.0588 0.0180
fect in the near-infrared band as well as the visible bands. 095 0.0371 0.0545 0.0528 0.3237
However, a second possible explanation is that it is more dif- D iff 0.0001 -0.0089 -0.0005 -0.0887 0.0246
ficult to get a good sampling of the ground when vegetation 111 0.0294 0.0465 0.0345 0.5309
is present, compared to when the soils are bare. Therefore, Diff 0.0059 -0.0077 0.0106 -0.0949 0.0298
the aircraft-based ground measurements of vegetated targets 175 0.0298 0.0541 0.0356 0.5126
could have a larger error than the soils targets. D iff
Ave/Diff/
- - - - -
-0.0012
0.0041
0.0182
-0.0109
0.0080
0.0230
0.0020
0.0045
-0.0587
0.0753
One reason it would be more noticable in the near-infra-
red band, and show up as a larger error in TM band 4, is be- *Moved f r o m vegetation section.

September 1996 PE&RS

 TABLE 6. COMPUTED REFLECTANCESGENERATED BY USING THE 5SD MODEL lower reflectances in the visible TM band 1, the additive
BASED ON A SIMULATEDATMOSPHERE AND DARKOBJECT OF 1PERCENT (MORAN correction is more important than the multiplicative correc-
ETAL., 1992), THE DIFFERENCESBEWEEN THE COMPUTED AND AIRCRAFT-BASED tion. Third, for high reflectances in TM band 1, the multipli-
GROUNDMEASURED VALUES,AND THE AVERAGEOF THE ABSOLUTE DIFFERENCE cative correction becomes more important; therefore, the
BETWEEN THESE VALUES.DOY, DAYOF YEAR, 1985-86; AVE, AVERAGE; DIFF, additive scattering component is not always the most im-
DIFFERENCE. portant correction for visible bands as indicated i n previous
studies (Turner et al., 1971; Slater, 1980). A similar relation
DOY TM1 TM2 TM3 TM4 AveIDiffl
exists for the near-infrared TM band 4, but the differences
204 are not as dramatic. These differences probably are part of
Diff the explanation of why the DOS model is actually the best
220 for very dark vegetated targets but generates unacceptable
Diff results for bright targets.
300 As an example, examine what happens to ground re-
Diff flectance targets having 10 and 35 percent reflectance val-
300* ues. In TM band 1the average additive scattering correction
Diff (Lhaze) for all seven dates is equal to about 7.8 percent re-
079 flectance. The average TAUz value of 0.70 generates a multi-
Diff plicative correction of about 3 percent reflectance for the 10
095 percent reflectance target; it took the additive corrected 7
Diff percent value to 10 percent (710.7 equals 10). However, for
111 the 35 percent reflectance target, the TM band 1additive
Diff correction is still 7.8 percent reflectance, but the multiplica-
175 tive correction is now 10.5 percent reflectance (24.510.7
equals 35). The additive correction amount remains the
same (7.8), while the multiplicative amount increases by a
factor of 3.5 (3 to 10.5). In the near-infrared TM band 4, the
Vegetation average additive scattering correction (Lhaze) for all seven
204 0.0640 0.0440 dates is equal to 1.75 percent reflectance. The average TAUz
Diff 0.0051 0.0171 value of 0.91 generates a multiplicative correction of about
220 0.0660 0.0500 0.9 percent reflectance for the 10 percent reflectance target
Diff 0.0069 0.0135 (9.1/0.91 equals 10). For the 35 percent reflectance target,
079 0.0520 0.0400 the additive correction is still 1.75 percent reflectance, but
Diff 0.0055 0.0098 the multiplicative correction is 3.15 percent reflectance
095 0.0730 0.0650 (31.8510.91 equals 35). The additive correction amount re-
Diff 0.0096 0.0117
111 0.0620 0.0430
Diff 0.0078 0.0191
TABLE7. COMPUTED REFLECTANCESGENERATED BY USING THE HBC MODEL
175 0.0730 0.0510
Diff 0.0140
BASED ON IN-SITU ATMOSPHERE MEASUREMENTS(MORANETAL., 1992), THE
0.0080 0.0174 -0.0023 0.0104
AveIDiffl DIFFERENCESBETWEEN THE COMPUTED AND AIRCRAFT-BASED GROUNDMEASURED
0.0135 0.0072 0.0148 0.0283 0.0159
VALUES, AND THE AVERAGEOF THE ABSOLUTE DIFFERENCES BETWEEN THESE
*Moved from vegetation section. VALUES. DOY, DAY OF YEAR, 1985-86; AVE, AVERAGE; DIFF, DIFFERENCE.

DOY (1985-86) TM1 TM2 TM4 AveIDiffl

however, on an overall average, the cosine function did a 204 0.2236
good job representing the TAUZ values for these data (0.80 av- Diff 0.0081 0.0066
erage cosine value versus 0.81 average TAUZ value). On the 220 0.2077
basis of TAU2 values in Table 3, it appears that the cosine Diff -0.0173 0.0211
model will generally undercorrect TM band 1and overcorrect 300 0.2335
TM band 4. Even though these data cover a semi-arid environ- Diff -0.0090 0.0035
ment, the cosine function and default TAU2 values computed 300* 0.2422
from these data may also satisfactorily approximate, to a h s t Diff -0.0015 0.0091
order, the kansmittance TAUZ values for continental and mar- 079 0.2080
itime atmospheres. Gilabert et al. (1994, Figure 3 on page Diff 0.0083 0.0079
2076) show the atmospheric transmittance values for eight 095 0.2102
continental and four maritime atmospheres corresponding to Diff 0.0063 0.0056
their Landsat TM data. The transmittance values for TM band 111 0.2522
1 for all 12 image dates range from 0.68 to 0.82, and continue Diff -0.0214 0.0159
to increase to approximately 0.95 at the longer near-infrared 175 0.2682
wavelength (p. 2074). In comparison, the transmittance values Diff 0.0020
- 0.0072
-
for TM band 1for the data used in this study (Table 3) range AveIDiffl 0.0092 0.0096
from 0.64 to 0.78; values also increase at the longer wave- Vegetation
lengths. The range of the cosine of the solar zenith angles for 204 0.0609 0.0367
these same data - 0.63 to 0.89 - implies that the improved Diff 0.0020 0.0098
image-based models may also be applicable under different 220 0.0594 0.0451
atmospheric conditions and nonarid environments. Further Diff 0.0003 0.0086
testing is required to see how well the relation actually does 079 0.0457 0.0343
hold up under different conditions. Diff -0.0008 0.0041
(2) Observe the relative amounts of correction contributed by
095 0.0757 0.0645
the additive versus multiplicative parameters. The TAUz val- Diff 0.0123 0.0112
ues i n Table 3 show the amount of correction needed for 111 0.0621 0.0420
each spectral band owing to multiplicative transmittance ef- Diff 0.0079 0,0181
fects. By analyzing and comparing the amounts of correc- 175 0.0702 0.0492
tion generated by the additive Lhaze and the multiplicative Diff 0.0116 0.0052 0.0156
TAUz parameters, we can see several things. First, both the - - -
AveIDiffl 0.0091 0.0048 0.0112
additive and multiplicative corrections are larger for the vis-
ible bands than for the near-infrared band. Second, for *Moved from vegetation section.

PE&RS September 1996 1033

 8. COMPUTED
TABLE REFLEcTANCES GENERATED BY USING THE IMPROVED IMAGE- sensor's characteristics change enough through time that the
BASEDCOSINEMODEL(COST),THE DIFFERENCES BETWEEN THE COMPUTED AND gain and offset values for any given image may not be accu-
GROUNDMEASURED
AIRCRAFT-BASED VALUES,AND THE AVERAGEOF THE ABSOLUTE rately known. This change can cause substantial errors i n
DIFFERENCEBETWEEN THESE VALUES. DOY, DAY OF YEAR, 1985-86; AVE. the conversion of DCS to at-satellite radiances, therefore giv-
AVERAGE;DIFF, DIFFERENCE. ing inaccurate reflectance values (Slater et al., 1987; Thome
- -
et ul., 1993; Moran et al., i n press). For multitemporal stud-
DOY ies, a different type of radiometric correction procedure can
204 0.1114 0.1630 be used to eliminate this problem. This procedure is men-
Diff -0.0091 -0.0054 tioned here only to present a solution to the potential gain
220 0.1112 0.1594 and offset accuracy problems encountered i n all the model-
Diff -0.0144 -0.0260 ing approaches discussed above. It is a hybrid method de-
300 0.0994 0.1216 veloped and used by Chavez and MacKinnon (1994) i n a
Diff 0.0053 0.0134 study dealing with automatic change detection of desert
300* 0.1343 0.1972 vegetation and involves making ground-reflectance measure-
Diff -0.0081 0.0025 ments of a dark and bright target during a satellite over-
079 0.1133 0.1637 flight. The in-situ field measurements are used to apply
Diff 0.0061 0.0098 what is called a brute force mapping of the recorded Dcs di-
095 0.0956 0.1438 rectly to ground reflectances. The brute force method, which
Diff -0.0036 -0.0060 does a linear mapping of DCS to ground reflectances on the
111 0.1257 0.1948 basis of ground readings, bypasses modeling corrections, so
Diff -0.0141 -0.0113 it eliminates dependence on the accuracy of the sensor's
175 0.1347 0.1955 gain and offset values. A disadvantage of the method is that
it requires in-situ field measurements; however, for multi-
Diff -0.0144 -0.0146 temporal studies such as theirs, the hybrid method mini-
AveIDiffl 0.0094 0.0112
mizes the need to be in the field during each satellite overf-
Vegetation light. The hybrid method radiometrically corrects historical
204 0.0573 0.0379 and new data to surface reflectances by slaving them onto
Diff -0.0016 0.0110 the reflectance master using a statistical matching proce-
220 0.0622 0.0481 dure. The sequence is (1) to radiometrically correct the one-
Diff 0.0031 0.0116 time-only in-situ field measurements of dark and bright
079 0.0558 0.0454 targets to the master image collected during the satellite ov-
Diff 0.0093 0.0152 erflight at the time of the ground measurements, then (2) to
095 0.0695 0.0673
Diff 0.0061 0.0140
111 0.0557 0.0413 TABLE9. COMPUTEO REFLECTANCESGENERATED BY USING THE IMPROVED IMAGE-
Diff 0.0015 0.0174 BASED DEFAULTTAUS MODEL,THE DIFFERENCES BETWEEN THE COMPUTED AND
175 0.0610 0.0402 AIRCRAR-BASED GROUNDMEASURED VALUES, AND THE AVERAGEOF THE ABSOLUTE
Diff -0.0040 DIFFERENCEBETWEEN THESE VALUES. DEFAULTTAUZ VALUES WERE0.70, 0.78,
Ave/Diff/ 0.0043 0.0127 0.85, AND 0.91. DOY, DAY OF YEAR, 1985-86; AVE, AVERAGE;DIFF,
"Moved from vegetation section. DIFFERENCE.
DOY TM1 TM2 TM3 TM4 AvelDiffl

mains the same (1.75), while the multiplicative amount Soils

again increases by 3.5 (0.9 to 3.15). Even though the in- 204 0.1238 0.1664
crease is by a factor of 3.5 for both spectral bands, the dif- Diff 0.0033 -0.0020
ference in the increase in terms of percent reflectance 220 0.1233 0.1622
between the 10 percent (dark) and 35 percent (bright) reflec- Diff -0.0023 -0.0232
tance targets is more dramatic in the visible TM band 1 than 300 0.0804 0.0902
in the near-infrared TM band 4 (7.5 and 2.25, respectively). Diff -0.0137 -0.0180
(3) Both of the new image-based models still ignore the effect 300* 0.1086 0.1464
of variations in downwelling which, according to Moran et Diff -0.0338 -0.0483
al. (1992, p. 178), can be considerable. However, the cosine 079 0.1049 0.1391
model generates results equivalent to the HBC model. There Diff -0.0023 -0.0148
are several possible reasons for this. First, downwelling irra- 095 0.0962 0.1327
diance, even though it can be equal to a substantial percent- Diff -0.0030 -0.0171
age of the ground radiance (up to 25 percent according to 111 0.1346 0.1914
Moran et al. (1992), p. 178), is in fact a relatively small per- Diff -0.0052 -0.0147
cent of the denominator in Equation 2. The first term i n the
denominator is much larger than the downwelling term;
therefore, the effect of downwelling is minimized. Also, in
175
Diff
Ave/Diff/
0.1531
0.0040
0.0085
-=
0.2039

0.0180
the model shown in Equation 2, the correction for the Vegetation
ground-to-satellite path (equal to EXP(-del*sec(TV))in 204 0.0637 0.0387
Moran et al. (1992)) was set to one because the cosine of the Diff 0.0048 0.0118
nadir view is equal to one; however, the values are actually 220 0.0690 0.0489
slightly less than one, i.e., i n the 0.85 to 0.95 range. This Diff 0.0099 0.0124
correction is a multiplicative coefficient in the denominator; 079 0.0517 0.0386
therefore, it tends to make the denominator smaller and the Diff 0.0257 0.0084
reflectance value larger, while the Edown downwelling ad- 095 0.0699 0.0621
ditive term makes the denominator larger and the reflec- Diff 0.0065 0.0088
tance value smaller. These two parameters, which were 111 0.0596 0.0406
both ignored in my image-based models, tend to work Diff 0.0054 0.0167
against each other, so part of the errors associated will can- 175 0.0694 0.0419
cel each other out.
(4) As stated earlier, all the radiometric-correction models have
the disadvantage that they are dependent on the accuracy of
the system's gain and offset values. One concern is that the *Moved from vegetation sectioq.

September 1996 PE&RS

 radiometrically correct the historical and new image data by measurements, Proc. 11th International Symposium on Remote
using a statistical histogram-matching procedure to "slave" Sensing of Environment, Ann Arbor, Michigan, pp. 731-755.
the images onto the corrected-reflectance master image. This Boas, M.L., 1966. Mathematical Methods in the Physical Sciences,
hybrid method makes the corrections independent of the John Wiley and Sons, Inc., New York, 778 p.
sensor's gain and offset values, which change with time, but
Chavez, P.S., Jr., 1975. Atmospheric, solar, and M.T.F. corrections
requires field measurements once. For historical and new
data sets, the procedure becomes entirely image based and for ERTS digital imagery, Proc. American Society of Photogram-
simple to apply. The need to make assumptions about over- metry Fall Conference, Phoenix, Arizona, p. 69.
all temporal changes or brightness distribution within the , 1988. An improved dark-object subtraction technique for at-
image, which is required for the histogram matching phase, mospheric scattering correction of multispectral data, Remote
must be done with care. For more details on the application Sensing of Environment, 24:459-479.
of the hybrid radiometric-correction procedure, see Chavez , 1989. Radiometric calibration of Landsat Thematic Mapper
and MacKinnon (1994). multispectral images, Photogrammetric Engineering b Remote
Sensing, 55(9):1285-1294.
Conclusions , 1992. Comparison of spatial variability in visible and near-
T w o entirely image-based radiometric-correction models that infrared spectral images, Photogrammetric Engineering b Re-
are not only easy a n d efficent to use, b u t also generate results mote Sensing, 58(7):957-964.
w i t h accuracies approximately equal t o those generated b y t h e Chavez, P.S., Jr., and D.K. MacKinnon, 1994. Automatic detection of
m o d e l using in-situ atmospheric measurements, have been vegetation changes i n the southwestern United States using re-
motely sensed images, Photogrammetric Engineering b Remote
presented. They are variations of t h e 110smodel w i t h t h e addi- Sensing, 60(5):571-583.
tion of a multiplicative correction for transmittance. T h e mul-
Curcio, J.A., 1961. Evaluation of atmospheric aerosol particle size
tiplicative transmittance values are derived b y using mostly
distribution from scattering measurement in the visible and in-
observed empirical relations that are straightforward and easy frared, Journal of the Optical Society of America, 51:548-551.
t o apply. T h e COST model works well for these images, w h i c h Fung, T., 1990. An assessment of TM imagery for land-cover change
represent a semi-aridlarid environment. However, both t h e co- detection, IEEE Transactions on Geoscience and Remote Sens-
sine of t h e solar zenith angle a n d t h e default TAUz values ing, 28(4):681-684.
seem t o approximate t h e transmittance values given b y Gila- Gilabert, M.A., C. Conese, and F. Maselli, 1994. An atmospheric cor-
bert et al. (1994) for a non-arid environment and different at- rection method for the automatic retrieval of surface reflectances
mospheric conditions. However, further testing i s n e e d e d n o t from TM image, International Journal of Remote Sensing, 15(10):
only for non-arid environments and different atmospheric con- 2065-2086.
ditions, b u t also for images having higher solar zenith angles Hall, D.K., A.T.C. Chang, and H. Siddalingaiah, 1988. Reflectances of
(greater than 55 degrees). A s i s t h e case w i t h t h e sun-angle glaciers as calculated using Landsat-5 Thematic Mapper data,
correction, t h e cosine-function correction u s e d for t h e multi- Remote Sensing of Environment, 25:311-321.
plicative transmittance effects b y t h e COST m o d e l m a y also Hall, D.K., A.T.C. Chang, J.L. Foster, C.S. Benson, and W.M. Koval-
over-correct a t higher zenith angles, s o t h e default TAUz m o d e l ick, 1989. Comparison of in situ and Landsat derived reflectance
will b e more appropriate for those images. of Alaskan glaciers, Remote Sensing of Environment, 28:23-31.
Off-nadir viewing corrections are n e e d e d for images col- Herman, B.M., and S.R. Browning, 1965. A numerical solution to the
lected b y either AVHRR o r off-nadir viewing SPOT. T h e good equation of radiative transfer, Journal of Atmospheric Science,
results generated b y t h e COST m o d e l for these data suggest 22:59-66.
that this s a m e correction function c o u l d b e considered for Holm, R.G., S.M. Moran, R.D. Jackson, P.N. Slater, B. Yuan, and S.F.
off-nadir viewing. T h e tables a n d scatter plots s h o w t h a t a l l Biggar, 1989. Surface reflectance factor retrieval from Thematic
m o d e l s t e n d t o overcorrect for very l o w reflectances (DOS t h e Mapper data, Remote Sensing of Environment, 27:47-57.
least) and t h a t t h e additive scattering correction is n o t al- Jackson, R.D., 1990. The MAC experiments, Remote Sensing of Envi-
w a y s t h e most important for t h e visible bands. T h e additive ronment, 32:77-79.
correction is m o r e important for t h e darker reflectances, a n d Jensen, J.R., and D.L. Toll, 1982. Detecting residential land-use de-
t h e multiplicative transmittance correction i s m o r e important velopment at the urban fringe, Photogrammetric Engineering b
for brighter reflectances. Remote Sensing, 48(4):629-643.
A s n o t e d b y M o r a n et al. (1992), these d a t a d o n o t cover Kneizys, F.X., E.P. Shettle, L.W. Abreu, and others, 1988. Users
t h e full range of reflectances (TM b a n d s 1 , 2, and 3 values are Guide to LOWTRAN 7, Report AFGL-TR-88-0177, Air Force
Cambridge Research Laboratories, Bedford, Massachusetts, 137
generally less t h a n 20 percent reflectance a n d TM 4 i s greater
t h a n 20 percent). If possible, future studies s h o u l d u s e tar- P.
Leprieur, C.E., J.M. Durand, and J.L. Peyron, 1988. Influence of to-
gets t h a t i n c l u d e more of t h e d y n a m i c range of reflectances pography on forest reflectance using Landsat Thematic Mapper
i n all TM bands. Differences i n accuracy between t h e visible and digital terrain data, Photogrammetric Engineering b Remote
and near-infrared b a n d s o r soils versus vegetation m a y actu- sensing, 54(4):491-496.
ally b e a bright versus dark situation. Markham, B.L., and J.L. Barker, 1986. Landsat MSS and TM post-cal-
ibration dynamic ranges, exatmospheric reflectances and at-sat-
Acknowledgments ellite temperatures, EOSAT Landsat Technical Notes, (1)3-8.
T h e a u t h o r t h a n k s Dr. S u s a n Moran for making available n o t Moran, M.S., 1986. The MAC experiment, Remote Sensing Newslet-
o n l y t h e d a t a p u b l i s h e d in Moran et al. (1992) b u t also u n - ter, University of Arizona, Tucson, (86):l-9.
p u b l i s h e d data u s e d t o derive those in t h e paper. Her sharing Moran, M.S., R.D. Jackson, G.F. Hart, and others, 1990. Obtaining
of data a n d information allowed t h e n e w m o d e l s t o b e better surface reflectance factors from atmospheric and view angle cor-
tested a n d c o m p a r e d w i t h current m o d e l s being w i d e l y used. rected SPOT-1 HRV data, Remote Sensing of Environment, 32:
Dr. M o r a n w a s also helpful w i t h an early review of this pa- 203-214.
per. I also t h a n k Dr. David MacKinnon for helpful discus- Moran, M.S., R.D. Jackson, P.N. Slater, and P.M. Teillet, 1992. Evalu-
sions a n d t h e early review of t h e paper. ation of simplified procedures for retrieval of land surface reflec-
tance factors from satellite sensor output, Remote Sensing of
Environment, 41:169-184.
References Moran, M.S., R.D. Jackson, T.R. Clarke, J. Qi, F. Cabot, K.J. Thome,
and P.N. Slater, in press. Reflectance factor retrieval from Land-
Ahern, F.J., D.G. Goodenough, S.C. Jain, V.K. Kao, and G. Rochon, sat TM and SPOT HRV data for bright and dark targets, submit-
1977. Use of clear lakes as standard reflectors for atmospheric ted to Remote Sensing of Environment.

PE&RS September 1996

Price, J.C., 1987. Calibration of satellite radiometers and the compar- of multispectral sensors, Remote Sensing of Environment, 2 2 : l l -
ison of vegetation indices, Remote Sensing of Environment, 21: 37.
15-27. Tanre, D.,C. Deroo, P. Duhaut, M. Herman, J.J. Morcrette, J. Perbos,
Robinove, C.J., P.S. Chavez, Jr., D. Gehring, and R. Holmgrem, 1981. and P.Y. Deschamps, 1990. Description of a computer code to
Arid land monitoring using Landsat albedo difference images, simulate the satellite signal in the solar spectrum: the 5 s code,
Remote Sensing of Environment, 11:133-156. International Journal of Remote Sensing, 11:659-668.
Sabins, F.F., Jr., 1978. Remote Sensing Principles and Interpretation, Thome, K.J., D.I. Gellman, R.J. Parada, S.F. Biggar, P.N. Slater, and
Freeman, San Francisco, 425 p. S.M. Moran, 1993. In-flight radiometric calibration of Landsat-5
Thematic Mapper from 1984 to present, Proc. Society of Photo-
Slater, P.N., 1980. Remote Sensing - Optics and Optical Systems,
Addison-Wesley, Reading, Massachusetts, 575 p. Optical Instrument Engineering Symposium, Orlando, Florida,
pp. 126-130.
, 1985. Radiometric considerations in remote sensing, Proc. of Turner, R.E., W.A. Malila, and R.F. Nalepha, 1971. Importance of at-
IEEE, 73(6):997-1011. mospheric scattering in remote sensing, Proceedings of 7th Inter-
, 1988. Radiometric calibration requirements and atmospheric national Symposium on Remote Sensing of Environment, Ann
correction, SPlE Proceedings, (924):144-150. Arbor, Michigan, pp. 1651-1697.
Slater, P.N., F.J. Doyle, N.L. Fritz, and R. Welch, 1983. Photographic Vincent, R.K., 1972. An ERTS Multispectral Scanner experiment for
systems for remote sensing, Manual of Remote Sensing, Second mapping iron compounds, Proc. 8th International Symposium
Edition, American Society of Photogrammetry, Vol. 1, Chap. 6, on Remote Sensing of Environment, Ann Arbor, Michigan, pp.
pp. 231-291. 1239-1247.
Slater, P.N., S.F. Biggar, R.G. Holm, and others, 1987. Reflectance- (Received 1 8 April 1985; revised and accepted 1 9 October 1995; re-
and radiance-based methods for the in-flight absolute calibration vised 16 November 1995)

Forthcoming Articles
Michael Abrams, Remo Bianchi, and Dave Pieri, Revised Mapping of N.G. Kardoulas, A.C. Bird, and A.I. Lawan, Geometric Correction
Lava Flows on Mount Etna, Sicily. of SPOT and Landsat Imagery: A Comparison of Map and
M. Aniya, H. Sato, R. Naruse, R Skvarca, and G. Casassa, The Use of GPS Derived Control Points.
Satellite and Airborne Imagery to Inventory Outlet Glaciers of the Steven T. Knick, John T Rotenberry, and Thomas J. Zarriello, Su-
Southern Patagonia Icefield, South America. pervised Classification of Landsat Thematic Mapper Imagery
Ling Bian and Eric West, GIs Modeling of Elk Calving Habitat in a in a Semi-Arid Rangeland by Nonparametric Discriminant
Priairie Environment with Statistics. Analysis.
Georges Blaha, Accuracy of Plates Calibrated by an Automatic Jacek Komorowski-Blaszczynski,Landform Characterization
Monocomparator. with Geographic Information Systems.
M. Les Bober, Duncan Wood, and Raymond A. McBride, Use of Digital Miklos Kovats, A Large-Scale Aerial Photographic Technique for
Image Analysis and GIs to Assess Regional Soil Compaction Risk. Measuring Tree Heights on Long-Term Forest Installations.
Gerardo Bocco and Hugo Riemann, Quality Assessment of Polygon Amnon Krupnik, Using Theoretical Intensity Values as Un-
Labeling. knowns in Multiple-Patch Least-Squares Matching.
Michel Boulianne, Clement Nolette, Jean-Paul Agnard, and Martin Kenneth C. McGwire, Cross-Validated Assessment of Geometric
Brindamour, Hemispherical Photographs Used for Mapping Con- Accuracy.
fined Spaces. Sunil Narumalani, John R. Jensen, Shan Burkhalter, John D.
Timothy L. Bowers and Lawrence C. Rowan, Remote Mineralogic and Althausen, and Halkard E. Mackey, Jr., Aquatic Macrophyte
Lithologic Mapping of the Ice River Alkaline Complex, British Modeling Using GIs and Logistic Multiple Regression.
Columbia, Canada, Using AVIRIS Data. Paul Pope, Ed Van Eeckhout, and Cheryl Rofer, Waste Site Char-
Stefan H. Cairns, Kenneth L. Dickson, and Samuel F. Atkinson, An acterization through Digital Analysis of Historical Aerial
Examination of Measuring Selected Water Quality Trophic Indica- Photographs.
tors with SPOT Satellite HRV Data. Elijah W. Ramsey III, Dal K. Chappell, and Dan G. Baldwin,
Ronald J. Duhaime, Peter V: August, and William R. Wright, Auto- AVHRR Imagery Used to Identify Hurricane Damage in a For-
mated Vegetation Mapping Using Digital Orthophotography. ested Wetland of Louisiana.
Christopher D. Elvidge, Kimberly E. Baugh, Eric A. Kihn, Herbert W. Tian-Yuan Shih, The Sign Permutation in the Rotation Matrix and
Kroehl, and Ethan R. Davis, Mapping City Lights with Nighttime the Formulation of Collinearity and Coplanarity Equations.
Data from the DMSP Operational Linescan System. R.D. Spencer, M.A. Green, and pH. Biggs, Integrating Eucalypt
Patricia G. Foschi and Deborah K. Smith, Detecting Subpixel Woody Forest Inventory and GIs in Western Australia.
Vegetation in Digital Imagery Using Two Artificial Intelligence M.D. Tomer, J.L. Anderson, and J.A. Lamb, Assessing Corn Yield
Approaches. and Nitrogen Uptake Variability with Digitized Aerial Infra-
Jay Gao and Stephen M. O'Leary, The Role of Spatial Resolution in red Photographs.
Quantifying Suspended Sediment Concentration from Airborne A.F? van Deventer, A.D. Ward, pH. Gowda, and J.G. Lyon, Using
Remotely Sensed Data. Thematic Mapper Data to Identify Contrasting Soil Plains
Greg G. Gaston, Peggy M. Bradley, Ted S. Vinson, and Tatayana F? and Tillage Practices.
Kolchugina, Forest Ecosystem Modeling in the Russian Far East Jianjun Wang, Gary J. Robinson, and Kevin White, A Fast Solu-
Using Vegetation and Land-Cover Regions Identified by Classifi- tion to Local Viewshed Computation Using Grid-Based Digi-
cation of GVI. tal Elevation Models.
Philip T Giles and Steven E. Franklin, Comparison of Derivative To- James D. Wickham, Robert I.! O'Neill, Kurt H. Ritters, Timothy G.
pographic Surfaces of a DEM Generated from Stereoscopic SPOT Wade, and K. Bruce Jones, Sensitivity of Selected Landscape
Images with Field Measurements. Pattern Metrics to Land-Cover Misclassification and Differ-
Clyde C. Goad and Ming Yang, A New Approach to Precision Air- ences in Land-Cover Composition.
borne GPS Positioning for Photogrammetry. Eric A. Williams and Dennis E. Jelinski, On Using the NOAA
Qizhong Guo and Norbert P. Psuty, Flood-Tide Deltaic Wetlands: De- AVHRR "Experimental Calibrated Biweekly Global Vegetation
tection of Their Sequential Spatial Evolution. Index. "
Joachim Hohle, Experience with the Production of Digital Orthophotos. Paul A. Wilson, Rule-Based Classification of Water in Landsat
Collin G. Homer, R. Douglas Ramsey, Thomas C. Edwards, Jr., and MSS Images Using the Variance Filter.
Allan Falconer, Landscape Cover-Type Mapping Modeling Using Zhangshi Yin and T.H. Lee Williams, Obtaining Spatial and Tem-
a Multi-Scene Thematic Mapper Mosaic. poral Vegetation Data from Landsat MSS and AVHRRINOAA
Pamela E. Jansma and Harold R. Lang, Applications of Spectral Satellite Images for a Hydrologic Model.
Stratigraphy to Upper Cretaceous and Tertiary Rocks in Southern Ding Yuan, A Simulation Comparison of Three Marginal Area
Mexico: Tertiary Graben Control on Volcanism. Estimators for Image Classification.

September 1996 PE&RS