Você está na página 1de 15

Remote Sensing of Environment 75 (2001) 335 ± 349

www.elsevier.com/locate/rse

A semianalytical approach to the calibration of AVIRIS data to


reflectance over water
application in a temperate estuary
John F. Mustard*, Matthew I. Staid, William J. Fripp
Department of Geological Sciences, Brown University, Providence, RI USA
Received 27 January 2000; accepted 8 August 2000

Abstract

Calibration of at-sensor radiance to reflectance is a critical step for the use of remotely sensed hyperspectral data of water to determine the
concentrations of optically active components. Because water has such a low reflectance, sources of radiance other than the water-leaving
radiance contribute significantly to the at-sensor radiance and vary with wavelength and spatially. While radiative transfer models have
improved significantly, there are still large uncertainties in prescribing the spatial and spectral properties of aerosol scattering. In addition,
errors in the sensor calibration of radiance can lead to additional uncertainty. An atmospheric correction method based entirely on scene
information, but not in situ data, is described. This approach accounts for nonuniform aerosol scattering, glint from the water surface, and
reflected skylight. Normalization by an estimate of the direct solar irradiance reaching the surface, also derived from scene information, then
provides an estimate of reflectance. This method is applied to AVIRIS data acquired over a temperate estuary. The resultant reflectance
estimates are shown to be consistent with field observations. While the method is empirical, it is based on physical principles that allow it to
be applied under a wide range of conditions. D 2001 Elsevier Science Inc. All rights reserved.

1. Introduction Analytical modeling of the inherent optical properties of


water require a precise determination of the surface reflec-
Remotely sensed data of water with hyperspectral sen- tance. However, the radiance measured by airborne or
sors have the potential to provide a wealth of information space-borne systems contains contributions from many
on the concentration of phytoplankton, dissolved organic sources other than water that must be removed (e.g., atmo-
matter, and suspended sediment as well as basic informa- spheric path radiance, Fresnel reflections off the water
tion on water clarity. Such information would provide great surface). Furthermore, the downwelling irradiance must be
benefits for understanding physical and biological pro- determined to ratio to the radiance measured at any given
cesses in coastal and estuarine systems. The advance that pixel. At blue wavelengths, the contributions from these
hyperspectral sensors offer over multispectral systems other sources may contribute in excess of 90% of the total
is that a continuous spectrum of the aquatic system signal, while measurable radiances farther into the infrared
is acquired. These added data permit more sophisticated, may be detected due to Fresnel reflectance. Under relatively
full spectral resolution models that allow greater precision clear atmospheric conditions, some of these contributions
and accuracy in determining the inherent optical proper- affect an image more or less uniformly (e.g., direct solar
ties of water and the concentration of absorbing species illumination). Other contributions are spatially variable,
(e.g., Bukata, Jerome, Kondratyev, & Pozdnyakov, 1995; ranging from broad variations due to asymmetries in atmo-
Lee, Carder, Mobley, Steward, & Patch, 1999; Roesler & spheric scattering to pixel by pixel variations due to reflec-
Perry, 1995). tions off wave facets.
Methods to convert the measured at-sensor radiances to
reflectance must account for these uniform and spatially
variable factors to be able to accurately apply analytical
* Corresponding author. models across a scene. A number of approaches have been
E-mail address: john_mustard@brown.edu (J.F. Mustard). developed ranging from fully analytical atmospheric mod-

0034-4257/00/$ ± see front matter D 2001 Elsevier Science Inc. All rights reserved.
PII: S 0 0 3 4 - 4 2 5 7 ( 0 0 ) 0 0 1 7 7 - 2
336 J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349

els (e.g., Green, Conel, & Roberts, 1993) to vicarious 2. Background


methods that incorporate field measurements concurrent
with remote data acquisitions and atmospheric models The at-sensor radiance measured by a remote sensing
(e.g., Carder et al., 1993) to fully empirical approaches system (Lsat) consists of several discrete sources (Fig. 1)
such as the empirical line method (e.g., Elvidge, 1988). generally described by
While analytical approaches that treat the atmosphere ÿ 
explicitly will ultimately be the best approach, it is not Lsat ˆ Lpath ‡ Tf Lglint ‡ Lw ‡ Lsky …1†
always possible to provide or solve for the complete range
of atmospheric and other variables required to derive an where Lpath is the path radiance or that portion of the
accurate reflectance for such a dark target as water. measured radiance due to atmospheric scattering, Lglint
Furthermore, small inaccuracies in sensor radiance calibra- and Lsky are due to direct solar and diffuse sky radiances,
tions may propagate through these calibrations and lead to respectively, reflected off the water surface, Lw is the
unacceptable errors in estimated reflectances. water leaving radiance, and Tf is the diffuse transmit-
In this article we develop a method for reflectance tance of the atmosphere between the surface and the
calibration of hyperspectral data over water that is based sensor (see Table 1).
primarily on scene information. It takes advantage of The key term in this equation for our purposes is Lw,
radiances measured of highly reflective, non-spectral since this contains the results of absorption and scattering
clouds and of water shaded by the direct beam of the in the water column. The other three terms are additive
sun where the radiance is dominated by path radiance and to the at-sensor radiance and need to be removed to
reflected diffuse sky irradiance. Reinersman, Carder, & isolate Lw. From Lw an estimate of the reflectance of the
Chen (1998) developed a method based on illuminated water can be derived.
and shadowed water that was used to provide an indepen- The ultimate goal in calibration is to estimate the
dent measurement of reflectance for satellite-sensor cali- reflectance of water as accurately as possible across the
bration. While rigorous, this method still required accurate relevant and important wavelength ranges (typically 0.35±
model inputs as well as adjustment of the sensor calibra- 0.8 mm). With reflectance, analytical and empirical models
tion to match in situ measurements of reflectance and that relate the concentrations of in-water constituents (e.g.,
radiance. The approach that we present here is rooted in phytoplankton, chlorophyll a, dissolved organic carbon, and
the fundamental interactions that are expected to occur suspended sediment) to reflectance can be developed and
during the transit of electromagnetic radiation through the applied. Reflectance is generally defined as the ratio of light
atmosphere to the surface and back up to the sensor. from a surface or target to that incident upon it. However,
However, since all the information required to calibrate there is a range of specific types of reflectance that depend
the measured radiances to reflectance are derived from the on how the target and source energies are measured.
scene itself, or with minimal in situ/vicarious data, the A number of analytical models for water reflectance are
approach can be portable to many different locations with formulated for volume reflectance (Rv) which is defined as
relative ease. We show that it is possible to derive a the ratio of upwelling irradiance to downwelling irradiance
reasonable estimate of the reflectance of water in an measured just below the air ±water interface. While desir-
estuarine and near-shore environment that is consistent able, remote sensing systems measure radiance, and to
with numerous in situ reflectance measurements. Further- convert water-leaving radiance to upwelling irradiance
more, the quality of the data and the calibration are requires modeling with attendant assumptions and uncer-
sufficient to permit modeling of reflectance using inherent tainties. The most commonly used definition of reflectance
optical properties. for water remote sensing is remote sensing reflectance (Rrs)

Fig. 1. Generalized energy pathways that contribute to the measured at sensor radiance of AVIRIS.
J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349 337

Table 1 the differences in path length from the scene center to the
List of Symbols and Their Definitions
edges, as well as differences in the phase function for
Term Definition Raleigh and aerosol scattering (Bukata et al., 1995). Never-
E0 Solar irradiance at the top of the atmosphere theless, it is expected that the variability in path radiance
Esky Downward diffuse sky irradiance will be a regular function of the observation and illumina-
Ed Downwelling irradiance at the top the surface
tion viewing angles.
Rv Volume reflectance
Rf Fresnel reflectance of water As mentioned above, Lsky refers to the radiance generated
Rrs Remote sensing reflectance by atmospheric scattering that is downwelled to the water
Rr Radiance reflectance surface and reflected from it. This radiance contribution is
Lsat At-sensor radiance the result of diffuse sky irradiance (e.g., no specific direction
Lw Water-leaving radiance
like the direct path of solar radiance), and the amount of sky
Lpath Path radiance
Radiance due to glint which is solar irradiance irradiance reflected from the water's surface is a relatively
Lglint specularly reflected from the water surface simple function of the real part (n') of the complex index of
Radiance due to reflected skylight which is sky irradiance refraction of water. Since n' is relatively constant across the
Lsky specularly reflected from the water surface wavelength region of interest, the contribution of Lsky to the
LsatB At-sensor radiance of a bright target
at-sensor radiance is typically modeled as the downwelling
LsatD At-sensor radiance of a dark target
Tq Downward diffuse transmission of the atmosphere sky irradiance multiplied by the Fresnel reflectivity of water
Tf Upward diffuse transmission of the atmosphere which is approximately 2% (Carder et al., 1993). Lsky is
attenuated by the diffuse transmittance of the atmosphere
between the surface and the satellite.
defined as the ratio of upwelling radiance to downwelling Lglint is due to the direct solar beam reflected off the
irradiance. Similarly to Rv, Rrs cannot be measured directly water surface. Since Lglint can overwhelm Lsat under unfa-
and the downwelling irradiance must be modeled using vorable viewing conditions, most remote sensing measure-
knowledge of the prevailing atmospheric conditions at the ments of water seek to constrain the viewing conditions
time of the measurement (e.g., Carder et al., 1993). Radi- such that the solar elevation is between 35° and 50° as well
ance reflectance (Rr) is defined as the ratio of the upwelling as for wind conditions of < 15 m/s. Contributions from glint
radiance to the radiance of a 100% reflective material may be further reduced by having the sensor flown toward
measured at the same viewing geometry. Rr may be more or away from the sun. However, it is not always possible to
easily determined from information within a remotely meet these conditions due to the characteristics of a field site
sensed data set and then converted to Rrs using an under- or scheduling. In addition, future orbiting instruments will
standing of the measurement conditions and the relation- have fixed orbital paths and thus will not be able to orient
ships between these measurements. along the solar plane. Thus a general ability to recognize
Returning to Eq. (1), in order to arrive at an estimate of and remove contributions from glint are required. While the
Rr it is necessary to isolate Lw as well as derive an estimate path radiance in a scene should exhibit a variable contribu-
of the radiance (L0) reflected from a 100% reflective tion that varies regularly with the viewing angles, glint is
lambertian surface for each pixel in the scene. While each expected to exhibit high variability on a pixel by pixel basis.
of the other terms in Eq. (1) is additive to the total measured
radiance, they are derived from different processes and have
different characteristics. Lpath is typically the largest con- 3. Methods of calibration
tribution to the measurement radiance in the short wave-
length portions of the visible, declining exponentially There are three basic approaches to the calibration of
toward longer wavelengths. Lpath is due to Raleigh scatter- hyperspectral remotely sensed data: atmospheric models,
ing from atmospheric molecules as well as Mie scattering empirical methods, and a model ±empirical hybrid. Atmo-
due to aerosols and the sum total of this radiance generally spheric models have been steadily improving and have been
follows a l ÿ 4 dependence. While Raleigh scattering is well employed with some success. The most robust approach is
understood and consistently present in remotely sensed data, that developed by Green et al. (1993), which is based on the
aerosol scattering is much more variable and highly depen- MODTRAN atmospheric model (Kneizys et al., 1987).
dent on the characteristics of the aerosols and their distribu- Among its strengths is that has been demonstrated to be
tion in the atmosphere. Because this is the dominant source effective in separating atmospheric water vapor from leaf
of radiance in the shorter wavelength region, small measure- water absorptions (Roberts, Green, & Adams, 1997). Green
ment, calibration, and model errors lead to large errors in the et al. (1998) note that the average overall agreement
determination of Lw (e.g., Bukata et al., 1995; Gastil & between MODTRAN calculations and AVIRIS measure-
Melack, 1998). ments for their well-characterized calibration site averages
While Lpath is typically the largest contribution to the > 96%. The residual uncertainty is a combination of model,
measured radiance over water, it is not spatially uniform target, and instrument errors. Nevertheless, for both the
across a remotely sensed scene. This variability is related to instrument and the models, the greatest uncertainty occurs
338 J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349

at the shorter wavelengths. Thus, for water applications, the length due to topography are not accommodated (Clark,
treatment of path and sky radiance is not as precise as Swayze, Keidebrecht, Green, & Goetz, 1995). In addition,
required for such applications. For example, Gastil and small measurement errors propagate into regression errors
Melack (1998) showed systematic differences between the that result in magnified reflectance errors when resolving
model estimates and field measurements that needed to be spectral differences among very dark targets like water
corrected in order to derive their accurate measurements of (Mustard & Staid, 1998).
water reflectance. Similarly, Carder et al. (1993) used a Hybrid methods use atmospheric models to accommo-
correction factor to match in situ radiance over water with date the main atmospheric transmission and scattering con-
that predicted from a Lowtran 7 model in order to achieve a tributions on a pixel by pixel basis, and then refine the
robust calibration. One of the most widely used atmospheric reflectance calculation using reference reflectance spectra of
models is ATREM (Gao, Keidebrecht, & Goetz, 1993). As a limited number of known targets (e.g., Clark et al., 1999).
with MODTRAN, however, the primary deficiencies are This approaches removes small artifacts caused by imper-
with the short wavelength regions critical to accurate fect knowledge of sensor wavelength position and resolu-
determination of water reflectance. While atmospheric mod- tion. While sufficient for observations of bright targets with
els are the preferred method for calibration, additional work relatively large spectral contrast, hybrid methods are still
is required for regular and systematic application. limited in the precision and accuracy required for resolving
Empirical methods take advantage of ground information spectral differences among water bodies.
to provide guidance for reflectance calibration. The most Another approach that has similarities to the hybrid
commonly used approach is termed the Empirical Line methods described above is the cloud-shadow approach
Method (Elvidge, 1988). Radiance data from two or more of Reinersman et al. (1998). Reinersman et al. (1998) used
regions of the surface that are fully resolved by the instru- the difference in radiance between fully illuminated water
ment and for which field reflectance spectra have been and water that is shadowed by clouds to constrain the
acquired are extracted. Provided that the reference sites water-leaving radiance for the illuminated water. By sub-
exhibit a range in albedo, the radiance measurements can tracting the radiance from the shadowed region and cor-
be regressed against the reference reflectance spectra to recting for some second-order effects, most of the
estimate surface reflectance. While it is simple in concept atmospheric effects are removed leaving radiance domi-
and has been used effectively, the principal drawbacks are nated by solar photons backscattered from beneath the
the requirements for well-characterized and spatially water surface. An estimate of Rrs is then provided by
resolved calibration sites in the field measured close in time dividing the water-leaving radiance by the downwelling
to the actual data collect and the fact that variations in path irradiance provided by the Lowtran 7 model. The authors

Fig. 2. Location map showing the AVIRIS flight line used in this analysis.
J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349 339

argue that this provides an independent estimate of water of a low-flying blimp to correct for path radiance and
reflectance that can be used in testing the calibration of the reflected skylight in multispectral video data. They
sensor. Carder, Carder, and Weigle (1997) used the shadow employed a ratio of the radiance measured in the shadow
to glint to achieve this result.
In the following section, we develop methods to deter-
mine the key radiance components in Eq. (1) as measured
by the Airborne Visible to InfraRed Imaging Spectrometer
(AVIRIS) (Green et al., 1998) and to use these radiances to
determine an estimate of water reflectance. These methods
are based on a basic understanding of the sources and
expected properties of the radiance terms, but are derived
primarily from the scene itself. The approach shares simila-
rities to that of Reinersman et al. (1998) in that the
difference in radiance between illuminated and shadowed
water are exploited to arrive and an atmospheric correction.
However, the approaches differ in that no modeling is
required for the methods presented here. It is also similar
to the methods advocated by Carder et al. (1997) except that
we employ radiances of clouds to determine the down-
welling irradiance while they employed glint measurements
to achieve the same result. In addition, Carder et al. (1997)
used 3-band video data while we apply this approach to
hyperspectral data.

4. Application to aviris data

The goal of calibration is to provide the best estimate of


reflectance from calibrated hyperspectral radiance data. In
this analysis we use AVIRIS data. Radiometrically corrected
data were provided by the AVIRIS Data Facility. Data were
acquired as 12-bit and converted to radiance in units of
microwatts per square centimeter per nanometer per ster-
adian, or mW/(cm2 nm sr), using inflight and ground
calibration files. AVIRIS radiometric calibration factors
are calculated by measuring the response of AVIRIS to an
integrating sphere (a known target illuminated by a known
light source). This calibration is reported to be accurate to
within 7%, absolute, over time while intraflight accuracy is
within 2% (Green et al., 1998). While this performance is
extremely good, a 7% absolute error could result in errors in
the measurement of Lw as high as 70% since 90% of the
measured radiance at short wavelengths is due to the atmo-
sphere. The data used in this analysis were obtained of

Fig. 3. At sensor radiance image of the AVIRIS flight line measured at a


wavelength of 0.51 mm. Land surfaces, clouds, and large boats have been
masked. The dominant patterns of light and dark observed in the water are
due to variable concentrations of backscattering material, likely suspended
sediment. Obvious patterns of glint are noted, and the slight brightening
observed over the western half of the image (most notable in the south) is
due to nonuniform path radiance. Box locations labeled 1 ± 4 indicate
where field data have been collected and correspond to: 1) Coastal Water,
2) Lower Estuary, 3) Middle Estuary, and 4) Upper Estuary. Box loca-
tions labeled A ± C indicate where transects from deep to shallow water
were extracted.
340 J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349

Narragansett Bay on 19 August 1997 at 11:01 on a flight


path generally oriented north ± south (Fig. 2). The solar
elevation angle was 52.0° and the solar azimuth was 141°.
The overall quality of the data is extremely good, with
signal-to-noise of between 500:1 and 1000:1 across the
wavelength region from the visible to the near-infrared
(Green et al., 1998). The field site imaged by the AVIRIS
instrument in the spectral channel 0.58 mm is shown in
Figure 3.

5. Identification and removal of nonuniform path


radiance and glint

Optimal view conditions for water targets include sun


elevations between 35° and 50° and orientation of the
sensor toward or away from the sun. Unfortunately, it is
not always possible to achieve these conditions and thus
methods to identify and correct for extraneous radiance
contributions are essential. Several such effects are
observed in the AVIRIS observations under consideration
here, illustrated in Figure 3. In the offshore region, glint off
ocean swells can clearly been seen along the eastern
boundary of the image, as well as nonuniform bright
regions in the estuary that are suggestive of glint off more
smooth water. Glint off waves and smooth water is not
observed west of the nadir observations (center of the
image). At wavelengths longer than 0.75 mm these patterns
of brightness persist and can in fact be observed throughout
Fig. 4. Radiance spectra extracted over the relatively homogenous waters at
the wavelength coverage of AVIRIS. A subtle overall
the southern end of the AVIRIS data. Both a) and b) show the same data
brightening is observed from east to west across the scene. plotted with different scales. Since the optical properties are homogenous,
This is much more evident at shorter wavelengths and is the differences observed are due to variable sources of radiance from glint
not observed for wavelengths greater than approximately and path radiance.
0.8 mm.
Both these effects can also be illustrated with radiance
spectra extracted from the scene, shown in Figure 4. These show approximately equivalent radiance in the visible, the
spectra were extracted from the southern part of the scene spectrum affected by glint shows a factor of 5± 10 greater
where the optical properties of the water exhibit little to no radiance at wavelengths greater than 1.5 mm.
variation. In Figure 4a, radiance extracted from areas that Path radiance observed by AVIRIS is expected to be
exhibited relatively greater amounts of path radiance, rela- nonuniform across the scene due to forward and backward
tively lesser amounts of path radiance, and relatively greater scattering in the atmosphere as well as the differences in
amounts of glint are plotted on a linear vertical scale of path length from the nadir view at the scene center to the
radiance, but a logarithmic horizontal scale of wavelength to 15° viewing angle at the scene edges. The variation in path
highlight differences in the short wavelength region of the radiance simply due to the path length differences is
spectrum. Note that the areas with glint and relatively high expected to be inversely proportional to the cos(e), where
path radiance have approximately equal radiance at 0.5 mm, e is the viewing angle (angle between the sensor and the
but different values at longer wavelengths. Likewise the surface normal). Variations due to forward and backward
spectrum with relatively lesser path radiance is very different scattering are described by the Raleigh and aerosol phase
from both the other spectra near 0.5 mm, but virtually identical functions (Gordon et al., 1983). While the phase functions
to the spectrum with a greater amount of path radiance at 1.0 can be complex, the Raleigh phase function is proportional
mm. If not taken into account on a pixel by pixel basis, these to cos2( g), where g is the scattering, or phase angle between
differences can translate into extremely large and erroneous the direction of direct solar radiation and the viewing angle
differences in reflectance. The effect of glint on measured of the sensor. Thus, to first order the nonuniform path
water radiance is best illustrated in Figure 4b. Here the same radiance should be a regular function of the phase angle
radiance data are plotted with a linear wavelength scale to 2.5 and the viewing angle, and should be proportional to
mm but a logarithmic radiance scale. While all the spectra cos2( g)/cos(e).
J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349 341

For the Narragansett Bay flight line, with a north ± south


flight path, solar elevation of 52°, and solar azimuth of 141°,
the phase angle varies from 45° along the eastern edge to
27° along the western edge. The ratio of cos2( g)/cos(e) for
these angles and the north ±south flight line is approxi-
mately linear as a function of the cross-track pixel, and thus
the contribution of path radiance should vary linearly across
the scene. This is in fact what is observed. Shown in Figure
5 is a profile of the 0.45 mm radiance extracted across the
scene over a homogeneous region of water which is not
affected by glint along with the ratio of cos2( g)/cos(e). The
ratio values are approximately linear across this phase angle
range, with a slight increase in the ratio toward small phase
angles. It is also evident that the radiance values show a Fig. 6. Radiance spectra derived from the AVIRIS scene used to correct for
similar trend, with an approximately linear relationship with nonuniform path radiance and glint.
phase angle across most of the range, with a slight increase
toward smaller phase angles. To first order, this trend is well
approximately by a linear function. way to the highest phase angle, the derived spectrum would
Since the magnitude of this contribution of path radiance be divided by 0.5 to scale it. The estimated nonuniform
is a regular function of phase angle, and since it is due to radiance for the radiance spectra shown in Figure 4 is shown
atmospheric scattering and thus not dependent on the in Figure 6.
reflectivity of the ground, it can be removed systematically This approach therefore uses the scene information to
from all the pixels in the scene. However, this additional derive the estimate of varying path radiance, but the
radiance will change as a function of wavelength. To magnitude of the correction is consistent with the known
accommodate the spectral variability, we derive an estimate characteristics of atmospheric scattering. It is not necessary
of the relative magnitude of the nonuniform path radiance that the locations used to derive the corrective spectrum
by extracting radiance spectra over water from two regions span the maximum range of observed scattering since it can
of the scene that have apparently uniform optical properties be scaled, as long as the variation in this relative path
not affected by glint, but that are observed under different radiance term is linear or can be modeled as a function of
phase angle conditions. The high phase angle spectrum is phase angle or cross-track pixel. More importantly, since the
subtracted from the low phase angle spectrum to derive an estimate of this path radiance is derived from the scene
estimate of the spectral properties of this path radiance. The itself, it is not subject to modeling errors, uncertainty in the
derived radiance is then scaled according to the relative paramterization of atmospheric scattering, or inaccuracies in
positions of the high and low phase angle spectra along the the absolute calibration of the AVIRIS radiance data. It does
path radiance gradient. For example, if the low phase angle depend, however, on the signal to noise of the sensor,
spectrum was extracted from the smallest angle and the high homogeneity of aerosols along the flight line, and the
phase angle spectrum was extracted from a location half assumption of homogeneity of optical properties between
the water bodies used to derive the path radiance correction.
The effect of glint on the measured radiance, as shown
in Figure 4, is to raise the overall radiance across the entire
wavelength range. Glint is due to direct reflection of solar
irradiance off the water surface and is controlled by the
relatively simple Fresnel interaction. As such, the reflected
radiance carries no spectral information due to volume
scattering in the water. In addition, since the refractive
index of water is relatively constant across visible to near-
infrared wavelengths, the magnitude of the reflected radi-
ance should be directly proportional to the incident solar
irradiance across the wavelength range. After calibration of
the radiance data to reflectance, glint should result in a
constant, additive increase in reflectance across the entire
Fig. 5. Profile of 0.45 mm radiance measured over water as a function of wavelength range.
phase angle (cross-track in Fig. 3) for a region with relatively homogenous It is a relatively simple exercise to remove glint by
optical properties compared with a function that describes the general
magnitude of aerosol and molecular scattering expected for the specific
identifying the magnitude of the contribution from a wave-
viewing geometry of the AVIRIS scene. Note that the radiance varies by as length region where the reflectance of water should be
much as 25% for these observed phase angles. effectively 0 (e.g., at 1.0 mm). This value would then be
342 J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349

subtracted from every wavelength in the calibrated reflec- of the image, it is not so evident that glint exists in the
tance data. Alternatively, a method similar to that employed estuarine regions. After correction for path radiance and
for the variable path radiance could be used. In this case, an glint, it is clear that there were significant regions of the bay
estimate of the spectral properties of glint can be derived by affected by glint. Note, however, that the glint correction
subtracting the radiance spectrum of a region unaffected by primarily affects regions to the east of the nadir point, which
glint from the radiance of a region which exhibits a large is expected for the viewing geometry of this scene.
contribution from glint. The resulting difference spectrum is
then the estimate of the spectral properties of glint. The
estimated contribution of glint for the radiance spectra
shown in Figure 4 is presented in Figure 6. In contrast to 6. Calibration to reflectance
the estimated path radiance spectrum, the estimated glint
spectrum mimics the general spectral properties of the solar The characterization and removal of the nonuniform
irradiance spectrum. The estimated contribution of glint to radiance contributions then permits approaches to be applied
each pixel can then be estimated by using a wavelength to estimate reflectance from the radiance measurements.
region where the spectral properties of water should effec- Estimation of reflectance over water bodies is complicated
tively be 0 (e.g., 1.0 mm). due to the relatively large contributions of path radiance
The visual effects of the removal of both nonuniform compared to water leaving radiance. Small errors in calcu-
path radiance and glint are shown in Figure 7 for the 0.58 lating and removing this component is not such a large
mm wavelength in a portion of the AVIRIS image near the problem with bright targets, but can lead to large inaccura-
mouth of Narragansett Bay. While the effects of glint are cies for dark targets like water. Furthermore, the contribu-
clearly observed in the wave patterns in the offshore regions tions from reflected sky irradiance compound these issues.

Fig. 7. An image of the measured at sensor 0.58 mm radiance (left) compared to the same data after correction for non-uniform path radiance and glint.
J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349 343

Assuming that the nonuniform contributions of path measurements from the scene where are being reflected
radiance and glint have been removed, the generic radiative from near the bottom of the atmosphere. If the dark target
transfer equation for radiance measured over water for nadir has a reflectance of 0%, then the radiance measured at the
views is given by sensor is simply due to path radiance and fresnel reflected
sky irradiance. In contrast, radiance measured over the
Rv   Esky Rf Tf bright target is controlled by Ed, which is dominated by
Lsat ˆ Tf T E0 cos ‡ Esky ‡ Lpath ‡ ; …2†
p p the direct solar beam. Furthermore, the attenuation of
where Rv is the volume reflectance of water, E0 is the solar irradiance in this wavelength region is controlled by
irradiance at the top of the atmosphere, Esky is the Raleigh and aerosol scattering. Since the absorptions are
downward diffuse sky irradiance, Rf is the fresnel relatively weak, the majority of the attenuated irradiance
reflectance of water, and Tq and Tf are the downward is converted to path radiance and thus Lpath is propor-
and upward diffuse transmissions of the atmosphere tional to (E0 ÿ E0Tq)/p. We propose that, to first order, the
respectively. The key objective we have is to estimate Rv. reduction in radiance from the surface to the sensor
Over water, Esky, Rf, and Lpath are convolved as an additive compared to the downwelling irradiance at the surface
term while Esky inside the brackets results in a relatively measured over a 100% reflective target is compensated by
minor contribution to the volume reflectance. While the input to the sensor from path radiance. In other words,
radiative transfer solutions to Eq. (2) have been used the at-sensor radiance measured over a 100% reflective
successfully to derive Rv, these approaches nevertheless lambertian target is approximately Ed/p. The consequences
include in situ data to correct for sensor and model errors of this assumption will be discussed later. Given these
(e.g., Carder et al., 1993; Reinersman et al., 1998) as well two approximations, the ratio shown in Eq. (3) can be
as require knowledge of the atmospheric conditions and simplified to
sea state. The method developed below instead utilizes LsatD Lpath ‡ Rf Esky =p
bright and dark components of the scene to estimate path  : …4†
LsatB Ed =p
radiance as well as the radiance expected from a 100%
reflective material to derive reflectance independent of in If these assumptions and approximations are correct, then
situ data and vicarious calibration. the ratio given in Eq. (4) will be a direct function of the path
Empirical methods for reflectance calibration have com- radiance in the scene and the reflected sky irradiance. The
monly used scene information to estimate the magnitude of radiance of a shadowed water as a dark target will also
the additive path radiance. These approaches typically use a contain a small contribution from volume reflectance,
dark object from the scene and assume that the reflectance is though this will be negligible beyond 750 nm.
close to zero. However, this approach cannot be used when Shown in Figure 8 are examples of ratios of radiance
investigating water, since water is commonly the darkest between a) water shadowed by a cloud and b) illuminated
material. Consider instead the ratio between a dark object water far from clouds, to a highly reflective, optically dense
such as water and a bright object of uniform spectral cloud. We examined a number of different regions from the
properties without significant absorptions. An ideal bright AVIRIS scenes and the examples shown are typical. Also
object would be 100% reflective surface. While such targets shown are the curves for the expected Raleigh (l ÿ 4.15) and
do not exist, cumulous clouds are relatively lambertian in aerosol scattering (l ÿ 1). The ratio spectra are plotted from
their scattering and exhibit no strong absorptions in the
visible-near infrared. The maximum reflectance of a cloud,
however, is approximately 65%. The consequences of this
assumption will be discussed later. An ideal dark object
would be a 0% reflective surface. Again, such objects do not
exist, but water that is shadowed by an optically thick cloud
provides a target that begins to approach this ideal. If we
assume that the volume reflectance of shadowed water is
0% and the reflectance of the cloud is 100%, then the ratio
of the two can be written as

LsatD Lpath ‡ Tf Rf Esky =p


ˆ ; …3†
LsatB Tf Ed =p ‡ Lpath

where LsatD is the at-sensor radiance of the dark object,


LsatB is the at-sensor radiance of the bright object, and Ed
Fig. 8. Ratios of the radiance measured over three different regions of the
is the downwelling irradiance (both direct and diffuse) at Narragansett Bay AVIRIS scene (fully illuminated water and two regions in
the surface. At this point we can ignore the irradiance at the shadows of clouds) compared to the general scattering curves expected
the top of the atmosphere since we are using radiance from molecular or Raleigh scattering (l ÿ 4.15) and aerosol scattering (l ÿ 1).
344 J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349

0.4 mm to 2.4 mm with the regions of strong atmospheric


absorptions removed. Data shortward of 0.4 mm are of poor
quality and will not be considered further. Note that the
illuminated water shows a greater ratio value than the
shadowed water in the 0.5 ±0.7 mm wavelength region. This
is due to volume reflectance in the illuminated water from
downwelling solar and diffuse irradiance. While the sha-
dowed water is not expected to exhibit a component of
volume reflectance due to direct solar irradiance, some
volume reflectance is expected due to diffuse irradiance,
most well expressed in the same wavelength region.
Note the general shape of the ratio spectra shown in
Figure 8 approximate that expected for the transmission of
Fig. 9. The same radiance ratio spectra shown in Figure 8 fit with an
the atmosphere due to the combined effects of Raleigh and
approximation to the path radiance scattering function. The fit is to the
aerosol scattering. Raleigh scattering is well known to lower convex hull of the radiance ratio spectra.
follow a l ÿ 4.15 dependence while aerosol scattering
exhibits a more diverse range of wavelength dependencies
ranging from l0 to l ÿ 2..5. The exact properties of the
aerosol scattering depend on the composition, size, dis- the data become more affected by lower signal to noise, and
tribution, and shape of the aerosol particles. The difficulty also the amount of path radiance has diminished signifi-
in specifying the exact nature of the aerosols and their cantly. Nevertheless, the shadowed water shows some
scattering leads to the largest errors in the modeling this component of volume reflectance of sky irradiance between
atmospheric component. 0.5 mm and 0.7 mm.
Since the numerators of the ratio spectra shown in This fit then provides a means to estimate the Raleigh
Figure 8 were obtained over water, the resultant ratios are and aerosol path radiance, as well as the contribution to the
expected to also contain a component of fresnel reflected measured radiance from reflected sky irradiance. This is
sky irradiance [Eq. (4)]. However, the sky irradiance is accomplished by multiplying the curve fit to the ratio
expected to have the same general properties of path spectra shown in Figure 9 by the original radiance measure-
radiance (e.g., spectral shape as a function of wavelength), ments of the bright, lambertian scattering cloud. At the same
and the fresnel reflectance of water is well approximated as time, as discussed above, we propose the cloud radiance
a constant across visible to near-infrared wavelengths spectrum provides a good first-order approximation of the
(Carder et al., 1993). Thus, to first order, the convolution direct solar irradiance at the surface. Thus, to calculate the
of path radiance and reflected sky irradiance is linear and radiance reflectance, Eq. (6) is applied:
would result in an enhancement of the curves shown in L…l†sat ÿL…l†path
Figure 8. Rr…l† ˆ ; …6†
L…l†0
Given that the ratio between low albedo water and highly
reflective clouds should be dominated by path radiance and where Rr is the estimated radiance reflectance, L * path is the
approximate the spectral properties of Raleigh and aerosol estimated path radiance, and L * 0 is the estimated down-
scattering, it should be possible to fit the resultant ratios welling irradiance reflected from a bright surface of uniform
with a function of the form: and high reflectance. A simple estimate of remote sensing
reflectance is achieved by dividing the Rr(l) by p. The
C1 C2 C3
Tp … l † ˆ 4:15
‡ 2‡ ; …5† resultant estimated values for L * path and L * 0 are shown in
l l l Figure 10. While this approach to estimating the critical path
where C1, C2, and C3 are constants to be determined radiance contributions over water is relatively simple, it is
interactively or analytically. Volume reflectance may based on the basic principles of radiative transfer.
contribute to the ratio across the wavelength region from The net result of these calculations is to derive calibration
0.4 mm to 0.7 mm. This is evident in comparing the ratio coefficients in the form of an offset term (L * path) and a gain
from illuminated water to the ratio from shadowed water. term (L * 0) to convert the measured radiance to an estimate
The illuminated water ratio spectrum shows a stronger of reflectance. This is analogous to the empirical line
departure from the overall transmission-like curve than the method where the bright object is assumed to be 100%
shadowed water. Thus, the fitting of Eq. (5) to the ratio reflective and nonspectral and the dark object 0% reflective
spectra should approximate the lower convex hull, and and also nonspectral. It differs in that the spectral properties
primarily at wavelengths longer than 0.7 mm. A representa- of the dark target are derived through the use of the dark/
tive fit of Eq. (5) to the ratio spectra is shown in Figure 9. bright spectral ratio which is fit with a function that con-
Note that the fit is best with the shadowed water and forms to the expected spectral characteristics of molecular
between wavelengths 0.4± 1.0 mm. At longer wavelengths, and aerosol scattering.
J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349 345

The assumption that the cloud reflects downwelling


irradiance with 100% efficiency results in errors in the
magnitude of the estimated reflectance but not in the
spectral shape. This can be demonstrated with radiance
spectra selected from a number of clouds in the AVIRIS
data that span a range of brightness (a function of their size
and optical depth), and two water spectra (one shadowed,
one fully illuminated). Procedures identical to those
described above were used to derive L * path and L * 0 for
each cloud spectrum using the same shadowed water
spectrum and applied to the illuminated water spectrum.
Thus separate pairs of L * path and L * 0 are derived for each
cloud spectrum. The results are shown in Figure 11. In
Figure 11a note that each cloud has a different radiance, and
in Figure 11b the derived reflectance is brighter for the
clouds with the lower radiance. However, upon normal-
ization to the reflectance at 0.56 mm, the derived reflectance
spectra have nearly identical values, indicating that the
spectral shapes are preserved. Thus, although the absolute
value of reflectance is dependent on the assumptions of the
cloud reflectance properties, because the spectral shape is
preserved analytical approaches that depend on the relative
reflectance properties (e.g., ratios, spectral parameters) will
not be affected by this assumption. Those that depend on the
absolute value of reflectance, however, will be affected
unless the reflectance properties of the cloud are accom- Fig. 11. a) radiance of three example clouds used to compute reflectance. b)
modated. Since the effect is scalar, the estimated spectra Reflectance resulting from assuming the clouds in a) are 100% reflective as
would only need to be divided by the reflectance of the well as reflectance normalized to 1. 0 at 0.56 mm. The normalized spectra
cloud to arrive at a more accurate reflectance. show that spectra shape is preserved by this calibration method.
Inaccuracies in the assumption that the attenuation of
radiance from the surface to the sensor is compensated by water. Alternatively, if the assumption results in an over-
path radiance between the surface and sensor [Eq. (4)] may estimate of the path radiance or if the path radiance is
affect the spectral shape primarily between 0.4 mm and 0.6 greater than the attenuation of radiance between the cloud
mm. If this results in an underestimate of the path radiance, and sensor, then the estimated reflectance will be too low in
or that the attenuation of radiance between the cloud and the the blue wavelengths. This will appear as increased absorp-
sensor is not adequately compensated by the path radiance, tion with spectral properties similar to color dissolved
then the resulting reflectance will be overestimate at blue organic matter or phytoplankton.
wavelengths. This will result in spectral properties that are
similar to strongly colored backscattering components in the
7. Resultant reflectance

The sequence of calibration steps developed in this


article were applied to the AVIRIS data for Narragansett
Bay. Ideally, an evaluation of the calibration would include
comparisons to reflectance spectra acquired in the field
concurrently with the acquisition of the AVIRIS data. This
was unfortunately not possible during the acquisition of the
1997 data. However, we have been systematically collecting
field reflectance spectra for four sites within the Bay over
the last year. The sites are located along a gradient in the
optical properties of the bay ranging from highly productive,
chlorophyll-rich waters of the upper bay to the less produc-
tive and optically distinct waters outside of the bay. While
the spectra do exhibit variability with time, each region
Fig. 10. Estimated solar and path radiance spectra used to determine exhibits a relatively consistent spectral shape that is distinct
reflectance through Eq. (6) from the other regions.
346 J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349

The in situ reflectance data were acquired with an Representative spectra from the four field sites are shown
Analytical Spectral Devices Personal Spectrometer II which in Figure 12 along with AVIRIS spectra from the same
covers the wavelength region 0.35 ±1.05 mm. Reflectance general regions extracted from the 1997 calibrated data set.
was calculated using the methods of Carder and Steward Spectra from the coastal waters, an offshore site, are typical
(1985). Briefly, the instrument integration time was set to of coastal or fjord type waters while the spectra from the
maximize the detector response of solar radiance reflected other three stations are typical of estuarine type waters (e.g.,
from a neutral gray target of 20% absolute reflectance. The Roesler & Perry, 1995). A good agreement is observed
instrument was adjusted for the dark current bias, and then between the field spectra and the remote acquired data. The
three main elements of the light field were sampled: diffuse general progression observed in the spectra properties from
radiance from the water (Lw), diffuse radiance from a region the coast to the upper estuary is toward stronger absorption
of the sky (Lsky), and diffuse radiance from the gray, neutral due to phytoplankton in the 0.35 ± 0.56 mm wavelength
target (Lgray). Reflectance is then calculated by Eq. (7): region and lower overall reflectance. In addition, the site
  farthest up the estuary exhibits enhanced reflectance
Lw ÿ 0:02  Lsky between 0.675 mm and 0.725 mm due to solar stimulated
Rr ˆ  Rgray …7†
Lgray chlorophyll fluorescence (Roesler & Perry, 1995). The
upper estuary and coastal waters exhibit the lowest reflec-
where Rr is the calculated radiance reflectance and Rgray is tance, while the lower estuary exhibits the highest reflec-
the absolute reflectance of the gray panel. The value of tance. The higher reflectance in the lower estuary is likely
0.02 is the average fresnel reflectance of water for a diffuse due to sediment that has become suspended due to scouring
sky irradiance. by tides. Thus the key diagnostic changes in reflectance

Fig. 12. Comparisons between field measurements of above-water reflectance and the reflectance estimated for the AVIRIS data using the methods developed
in this article. The field data and AVIRIS data were not obtained at the same time, although the field spectra are representative of the spectral properties of the
four regions used in the comparison. See text for details.
J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349 347

observed in the field spectra are faithfully reproduced in the


AVIRIS reflectance spectra.
Ideally, an exhaustive sampling of the waters of Narra-
gansett Bay with in situ sensors concurrent with the
AVIRIS overflight would be used to assess the accuracy
of the calibration. Nevertheless, our field observations
indicate that the optical properties of Narragansett Bay
show a consistent gradient through time from the phyto-
plankton-rich waters of the upper estuary to the less-rich
waters of the coastal ocean. This is consistent with
intensive study of the bay by Kremmer and Nixon
(1978) and recent systematic observations of its physical
and biological properties (Berman & Deacutis, 1999).
Therefore, the relatively simple approach to calibration
based completely on scene-derived information, but not
dependent on in situ or ground truth observations, is
capable of providing an accurate first-order estimate of
reflectance of water.

8. Diversity in optical properties

While the four systematically sampled sites provide a


good measure of the accuracy of the AVIRIS reflectance
calibration, it is also worthwhile examining the diversity in
types of water reflectance exhibited in these data and the
types of information that might be derived. Water reflec-
tance is controlled by the concentrations of phytoplankton,
colored dissolved organic matter, suspended sediment, sub- Fig. 13. Examples of variations exhibited in the calibrated AVIRIS scene.
The plot on top shows reflectance variations caused in part by differences in
strate reflectance in shallow water, and the fundamental the concentration of suspended sediment while the plot below shows
absorption and scattering properties of water itself. Many variations caused by differences in phytoplankton concentration. The large
researchers have examined the complex problem of decon- peak between 0.675 mm and 0.75 mm is due to fluorescence.
volving the contributions of these components to the mea-
sured reflectance and a number of analytical models exist
(e.g., Bukata et al., 1995; Lee, Carder, Peacock, Davis, & Narragansett Bay in Figure 7. The spectra shown in Figure
Mueller, 1996; Roesler & Perry, 1995; Zaneveld, 1995). In 13 are from a transect that crosses this plume. The effects of
general, the reflectance is a ratio of the total contributions increasing concentrations of suspended sediment are to
from backscattering to the total contributions from absorp- increase the backscattering from water and thus increase
tion. Deconvolution of a reflectance spectrum to the con- the overall reflectance, but nonlinearly. For example, the
centration of absorbing and scattering constituents is thus a differences in reflectance at 0.55 mm are greater than at 0.7
nonlinear problem and the subject of current research. The mm, and this is a well-known effect of the rapid changes in
range of optical properties exhibited in Narragansett Bay the absorption properties of water between 0.6 mm and 0.9
provide an excellent opportunity to test models under a mm (e.g., Pope & Fry, 1997; Reinersman et al., 1998). The
variety of conditions in a small area. effects of large concentrations of phytoplankton are illu-
As discussed above, variations in sediment and phyto- strated in Figure 13 where spectra from eutrophic fresh
plankton concentrations control the majority of the differ- water and brackish ponds are compared to typical upper
ences between the spectra shown in Figure 12. However, estuarine water. At typical estuarine concentrations, increas-
this limited sampling fails to capture the diversity of spectral ing the amount of phytoplankton primarily results in
water types observed in the AVIRIS data. Shown in Figure increased absorption. Also, between 0.625 mm and 0.725
13 are examples of spectra that exhibit more extreme mm there is a variable component of the measured reflec-
variations in phytoplankton and suspended sediment con- tance due to fluorescence. However, as the concentrations
centration. In Narragansett Bay erosion and suspension of increase to very high concentrations such as in eutrophic
bottom sediment due to tides are a common phenomenon in ponds, the phytoplankton contribute to backscattering, as
the middle reaches of the estuary. For example, a distinct well as very high amounts of fluorescence. Such spectra are
plume of estuarine water with suspended sediment entrained very unique and indicative of environmental problems that
can be observed at the mouth of the western arm of could be identified remotely.
348 J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349

The last example illustrates the effects of bottom reflec- shown in Figure 13 for varying concentrations of suspended
tion on the observed water reflectances. Bottom effects are sediment. This is to be expected since analytically the
well known to impact water observations, and a number of observed reflectance is directly proportional to the bottom
analytical models exist to account for these effects (e.g., reflectance modified by the optical depth properties of the
Carder et al., 1993; Lee et al., 1999). Calibrated spectra overlying water column (Carder et al., 1993) while the
from a diverse optical environment like Narragansett Bay presence of suspended sediment in the water will increase
would provide a good test of such models. Several transects the amount of backscattering. The similarity of the trends,
from deep to shallow water in a range of environments are however, raises the question whether separation of bottom
shown in Figure 14. The dominant effect is similar to that effects from suspended sediment will be possible in com-
plex estuarine environments.

9. Conclusions

While remote sensing of ocean water has progressed


rapidly over the last decade, our understanding of coastal
and inland waters still poses may important challenges.
Global observations from the SeaWiFS program has demon-
strated that multispectral observations can be used to accu-
rately estimate phytoplankton concentrations over Case I
waters. However, for Case II waters, it is likely that high
spectral resolution sensors will be needed to accurately
deconvolve the complex convolution of signals measured
remotely. Deconvolution approaches require an accurate
estimate of reflectance.
Because water is such a dark substance, it is important
that all extraneous sources of radiance not related to volume
reflectance of water be understood, characterized, and
removed. For the particular viewing geometries and geo-
graphic location of this study, large extraneous sources of
radiance were contributed from viewing-angle-dependent
path radiance and glint. The nonuniform path radiance
exhibited systematic variations with viewing angle expected
for the phase-angle dependence of path radiance, and thus
could be readily removed. Glint, on the other hand, is highly
variable and dependent on wind speed, the length and height
of waves, and viewing geometry. Nevertheless, the spectral
properties of glint are simply a function of the reflected solar
irradiance, and the magnitude of this contribution can be
identified and separated from path radiance using the
extended wavelength range of the AVIRIS sensor. While
the specific properties and distributions of the nonuniform
radiance and glint were specific to the scene analyzed, the
methods used were based on physical principles and can be
generalized and extended to other observations under dif-
ferent conditions.
We demonstrated here that accurate estimates of reflec-
tance can be derived from hyperspectral data over water
targets using only scene information. This is potentially
important since it obviates the need to exactly match model
and sensor radiance and allows the full precision of a sensor
to be exploited. The most problematic aspects of analytical
Fig. 14. Examples of variations exhibited in the calibrated AVIRIS scene
approaches to reflectance calibration for water are inaccura-
due primarily to decreasing the water depth. In each example, the optical
properties of the water have not changed, only the depth of the water. The
cies in the absolute measurement of radiance by the sensor
three examples illustrate the same effects but with water of differing optical and uncertainties in spectral properties of aerosols. The
properties overlying the bottom. approach developed here exploits the fact that the radiance
J.F. Mustard et al. / Remote Sensing of Environment 75 (2001) 335±349 349

over dark objects is dominated by path radiance while over Clark, R. N., Swayze, G. A., Keidebrecht, K., Green, R. O., & Goetz, A. F.
H. (1995), Calibration to surface reflectance of terrestrial imaging spec-
bright targets it is dominated by solar irradiance. Thus the
trometer data: comparison of methods. In Summaries of the Fifth JPL
ratio of dark to bright provides a first-order estimate of the Airborne Earth Science Workshop, JPL Publ. 95 ± 1, Jet Propulsion
transmission due to the combined effects of molecules and Laboratory, Pasadena, CA, Vol. 1, pp. 41 ± 42.
aerosols in the atmosphere. This itself may provide a benefit Clark, R. N., Swayze, G. A., King, T. V. V., et al. (1999), Surface reflectance
in determining aerosol properties analytically. calibration of terrestrial imaging spectroscopy data: a tutorial using
Although the final calculation of reflectance took advan- AVIRIS. In Summaries of the Ninth JPL Airborne Earth Science Work-
shop, CA Publ. 99 ± 1, Jet Propulsion Laboratory, Pasadena, CA, Vol. 1, pp.
tage of optically thick cumulous clouds present in the scene, 41 ± 42. http://makalu.jpl.nasa.gov/docs/workshops/98_docs/toc.htm.
other bright targets with neutral spectral properties can play Elvidge, C. D. (1988), Examination of the spectral features of vegetation in
a similar role, especially if the spectral properties of the 1987 AVIRIS data. In Proceedings of the Second Airborne Visible/In-
targets are known or can be estimated using suitable analy- frared Imaging Spectrometer (AVIRIS) Workshop, JPL Publ. 90 ± 54, Jet
tical calibration approaches (e.g., Gao et al., 1993; Green et Propulsion Laboratory, Pasadena, CA, pp. 97 ± 101.
Gao, B.-C., Keidebrecht, K. B., & Goetz, A. F. H. (1993). Derivation of
al., 1993). Errors introduced by this approach are possible in scaled surface reflectances from AVIRIS data. Remote Sensing of En-
two catagories: spectral and nonspectral. The assumption vironment, 44, 165 ± 178.
that the attenuation of radiance from the cloud top to the Gastil, M., & Melack, J. M. (1998), Improved atmospheric correction for
sensor is compensated by the addition of path radiance to the AVIRIS spectra from inland waters. In Summaries of the Eighth Annual
sensor can result in spectral errors. These errors will have a JPL Airborne Earth Science Workshop, JPL Publ. 98 ± 1, Jet Propulsion
Laboratory, Pasadena, CA, Vol. 1, pp. 41 ± 42. http://makalu.jpl.nasa.-
form that increasing or decreasing reflectance as a function gov/docs/workshops/98_docs/toc.htm.
of wavelength and will appear as increased absorption or Gordon, H. R., Clark, D. K., Brown, J. W., Brown, O. B., Evans, R. H., &
scattering in analytical models. The use of clouds for the Broendow, W. W. (1983). Phytoplankton pigment concentrations in the
bright targets assumes that they are 100% reflective. This Middle Atlantic Bight: Comparison of ship determinations and CZCS
results in an error in reflectance that is scalar. In other words, estimates. Applied Optics, 22, 20 ± 36.
Green, R. O., Conel, J. E., & Roberts, D. A., (1993), Estimation of aerosol
the estimated reflectance is related to the absolute reflec- optical depth and additional atmospheric parameters for the calculation
tance by a multiplicative factor that is constant as a function of the reflectance from radiance measured by the Airborne Visible/
of wavelength and is thus nonspectral. Nevertheless, analy- Infrared Imaging Spectrometer. In Summaries of the Fourth Annual
tical methods that rely on spectral shape (e.g. spectral ratios) JPL Airborne Geoscience Workshop, JPL Publ. 93 ± 26, Jet Propulsion
Laboratory, Pasadena, CA, Vol. 1, pp. 73 ± 76.
will be unaffected by this error.
Green, R. O., Eastwood, M. L., Satrure, C. M., et al (1998). Imaging
spectroscopy and the Airborne Visible/Infrared Imaging Spectrometer
(AVIRIS). Remote Sensing of Environment, 65, 227 ± 248.
Acknowledgments Kneizys, F. X., Shettle, E. P., Anderson, G. P., et al. (1987), Atmospheric
Transmittance/Radiance; Computer Code LOWTRAN 7, U.S. Air
Force Geophysics Laboratory, Hanscom AFB, MA.
Support from NASA's Commercial Remote Sensing Kremer, J. N., & Nixon, S. W. (1978), A Coastal Marine Ecosystem,
program (NAG13-39) is gratefully acknowledged as well as Springer-Verlag, New York, 217 pp.
the expert operation and calibration of the AVIRIS sensor by Lee, Z. P., Carder, K. L., Peacock, T. G., Davis, C. O., & Mueller, J. L.
the AVIRIS experiment team. In addition, we wish to thank (1996). Method to derive ocean absorption coefficients from remote-
sensing reflectance. Applied Optics, 35, 5721 ± 5732.
Jim Yoder, Warren Prell, Steve Clemens, and Darryl Keith Lee, Z. P., Carder, K. L., Mobley, C. D., Steward, R. G., & Patch, J. S.
for logistical support and thoughtful discussions and the (1999). Hyperspectral remote sensing for shallow waters: 2. Deriving
insightful comments of an anonymous reviewer. bottom depths and water properties by optimization. Applied Optics, 38,
3831 ± 3843.
Mustard, J. F., & Staid, M. L. (1998), Inverse modeling of reflectance to
obtain water constituent abundances: application to AVIRIS data of a
References temperate estuary. In Proc. 5th International Conference on Remote
Sensing for Marine and Coastal Environments, ERIM International,
Berman, M. S., & Deacutis. C. (1999), A cooperative study of Narragansett Ann Arbor, MI, Vol. 1, pp. 29 ± 35.
Bay, Rhode Island, USA, Abstract, American Society of Limnology and Pope, R., & Fry, E. (1997). Absorption spectrum (380 ± 700 nm) of
Oceanography. Aquatic Science Meeting. pure waters: II. Integrating cavity measurements. Applied Optics, 36,
Bukata, R. P., Jerome, J. H., Kondratyev, K. Y., & Pozdnyakov, D. V. 8710 ± 8723.
(1995), Optical Properties and Remote Sensing of Inland and Coastal Reinersman, P. N., Carder, K. L., & Chen, F.-I.R. (1998). Satellite-sensor
Waters, CRC Press, Boca Raton, FL, 362 pp. calibration verification using the cloud-shadow method. Applied Optics,
Carder, K. L., & Steward, R. G. (1985). A remote-sensing reflectance 37, 5541 ± 5549.
model of a red tide dinoflagellate of west Florida. Limnology and Roberts, D. A., Green, R. O., & Adams, J. B. (1997). Temporal and spatial
Oceanography, 30, 286 ± 298. patterns in vegetation and atmospheric properties from AVIRIS. Remote
Carder, K. L., Reinersman, P., Chen, R. F., Muller-Karger, F., Davis, C. O., Sensing of Envionment, 62, 223 ± 240.
& Hamilton, M. (1993). AVIRIS calibration and application in coastal Roesler, C. S., & Perry, M. J. (1995). In situ phytoplankton absorption,
oceanic environments. Remote Sensing of Environment, 44, 205 ± 216. fluorescence emission, and particulate backscatter determined from re-
Carder, M., Carder, K., & Weigle, B. L. (1997), Skylight correction of flectance. Journal of Geophysical Research, 100, 13,279 ± 13,294.
video images: an airship shadow approach. In Proc. 4th International Zaneveld, J. R. V. (1995). A theoretical derivation of the dependence of the
Conference on Remote Sensing for Marine and Coastal Environments, remotely sensed reflectance of the ocean on the inherent optical proper-
ERIM International, Ann Arbor, MI, Vol. II, pp. 631 ± 639, 1997. ties. Journal of Geophysical Research, 100, 13,135 ± 13,142.

Você também pode gostar