Spectral parameters of inherent optical property models:

method for satellite retrieval by matrix inversion

of an oceanic radiance model

Frank E. Hoge and Paul E. Lyon

Inherent optical property ~IOP! spectral models for the phytoplankton absorption coefficient, chro-
mophoric dissolved organic matter ~CDOM! absorption coefficient, and total constituent backscattering
~TCB! coefficient are linear in the reference wavelength IOP and nonlinear in the spectral parameters.
For example, the CDOM absorption coefficient IOP aCDOM~li ! 5 aCDOM~lref!exp@2S~li 2 lref!# is linear in
aCDOM~lref! and nonlinear in S. Upon linearization by Taylor’s series expansion, it is shown that
spectral model parameters, such as S, can be concurrently accommodated within the same conventional
linear matrix formalism used to retrieve the reference wavelength IOP’s. Iteration is used to adjust for
errors caused by truncation of the Taylor’s series expansion. Employing an iterative linear matrix
inversion of a water-leaving radiance model, computer simulations using synthetic data suggest that ~a!
no instabilities or singularities are introduced by the linearization and subsequent matrix inversion
procedures, ~b! convergence to the correct value can be expected only if starting values for a model
parameter are within certain specific ranges, ~c! accurate retrievals of the CDOM slope S ~or the
phytoplankton Gaussian width g! are generally reached in 3–20 iterations, ~d! iterative retrieval of the
exponent n of the TCB wavelength ratio spectral model is not recommended because the starting values
must be within ;65% of the correct value to achieve accurate convergence, and ~e! concurrent retrieval
of S and g ~simultaneously with the phytoplankton, CDOM, and TCB coefficient IOP’s! can be accom-
plished in a 5 3 5 iterative matrix inversion if the starting values for S and g are carefully chosen to be
slightly higher than the expected final retrieved values. © 1999 Optical Society of America
OCIS codes: 010.4450, 280.0280, 300.6550, 030.5620.

1. Introduction cifically, these three IOP’s, aph~lg!, ad~ld!, and bbt~lb!,

It has been shown that it is feasible to simultaneously can be readily retrieved from
retrieve the phytoplankton absorption coefficient, the
chromophoric dissolved organic matter ~CDOM!yde-
aph~lg!exp@2~li 2 lg!2y2g2# 1 ad~ld!exp@2S~li 2 ld!#
tritus absorption coefficient, and the total constituent
backscattering ~TCB! coefficient inherent optical 1 bbt~lb!~lbyli !nv~li ! 5 2aw~li ! 2 bbw~li !v~li ! (1)
properties ~IOP’s! from water-leaving radiances by
direct matrix inversion of a radiance model.1 Spe-
by linear matrix inversion.1 ~See Ref. 1 for nomen-
clature.! The IOP spectral model parameters g, S,
and n vary spatially and temporally in the global
oceans. Thus ~a! assigning constant values or ~b!
empirically derived values for these parameters can
lead to significant errors in the retrieved IOP’s. A
more general approach is to retrieve from the water-
F. E. Hoge is with the NASA Goddard Space Flight Center,
leaving radiances one or more of the g, S, and n
Wallops Flight Facility, Wallops Island, Virginia 23337. P. E.
Lyon is with EG&G Services, Incorporated, Wallops Flight Facil-
parameters simultaneously with the IOP’s.
ity, Wallops Island, Virginia 23337.
Received 17 August 1998; revised manuscript received 7 Decem-
2. Theory
ber 1998.
0003-6935y99y091657-06$15.00y0 Retrieval of the IOP spectral model parameters is
© 1999 Optical Society of America accomplished by linearizing the models by Taylor’s

20 March 1999 y Vol. 38, No. 9 y APPLIED OPTICS 1657

series expansion. Recall that the general Taylor’s ues n0, S0, and g0. Obviously, six sensor bands are
series for f ~x, y! about the point x0, y0 is2 required to provide a consistent solution because the
data-model matrix D is a 6 3 6 square matrix. ~If
f ~x, y! 5 f ~x0, y0! 1 ~]fy]x!x0,y0~x 2 x0! one wishes to retrieve only one spectral parameter
1 ~]fy]y!x0,y0~ y 2 y0! together with the three IOP’s, then only that IOP
spectral model would be expanded in a Taylor’s se-
1 1y2$~]2fy]x2!x0,y0~x 2 x0!2 ries. Of course, in this case the resulting inversion
1 2]2fy]x]y~x 2 x0!~ y 2 y0! would have a 4 3 4 D matrix.! In principle, any
number of spectral model parameters and their as-
1 ~]2fy]y2!x0,y0~ y 2 y0!2% 1 . . . . (2) sociated IOP’s can be accommodated within the lin-
ear matrix retrieval formalism. By induction one
If x0, y0 is a close approximation to the required so- can have K absorbers and backscatterers with their K
lution, then ~ x 2 x0! and ~ y 2 y0! are small, their spectral model parameters. Then the data-model
squares and products are negligible, and the higher- matrix D is a 2K by 2K. Unfortunately, as K in-
order terms can be ignored. Expanding the TCB creases, the IOP spectral models by necessity become
model bbt~lb!~lbyli !n in a Taylor’s series about the more complex because they must now fit through K
point bbt~lb! 5 bbt0~lb! and n 5 n0 yields wavelengths.
bbt~lb!~lbyli !n 5 ~lbyli !n0bbt~lb! 3. Results
1 bbt0~lb!~lbyli !n0~n 2 n0!ln~lbyli ! 1 . . . , (3)
A. Generation of the Simulated Water-Leaving Radiances
retaining only linear terms. Similarily, the Taylor’s A radiance model1 was used to produce the spectral
series expansion for the CDOMydetritus model radiances needed to test the iterative inversion meth-
ad~ld!exp@2S~li 2 ld!# about ad0~ld! and S0 is odology. A set of 10,000 water-leaving radiances
were calculated for a wide range of phytoplankton
ad~ld!exp@2S~li 2 ld!# 5 ad~ld!exp@2S0~li 2 ld!#
absorption, CDOM absorption, and TCB coefficients,
2 ad0~ld!exp@2S0~li 2 ld!#~li 2 ld!~S 2 S0! 1 . . . , typical of concentrations found in the Middle Atlantic
(4) Bight during spring conditions. ~The Middle Atlan-
tic Bight is representative of much of the world’s
and the Taylor expansion of the phytoplankton ab- oceans as it contains coastal, shelf, slope, Gulf
sorption coefficient model about aph~lg! 5 aph0~lg! Stream, and Sargasso Sea water masses.! The phy-
and g 5 g0 is toplankton Gaussian spectral width, CDOM spectral
slope, and the TCB wavelength ratio exponent were
aph~lg!exp@2~li 2 lg!2y2g2# set to constant values that are typical for each spec-
5 aph~lg!exp@2~li 2 lg0!2y2g02# tral model. Specifically, during computation of the
simulated radiances, values for the spectral model
1 aph0~lg!~li 2 lg!2~ g023! parameters were fixed at g 5 85 nm, S 5 0.017 nm21,
3 ~ g 2 g0!exp@2~li 2 lg0!2y2g02# 1 . . . . (5) and n 5 3.5.
B. Spectral Model Parameter Retrieval Results
Substitution of Eqs. ~3!, ~4!, and ~5! into Eq. ~1! yields We chose sensor bands near those of satellite sen-
the matrix equation sors such as the sea-viewing wide field-of-view sen-
sor, ocean color and temperature scanner, and the
aph~lg!exp@2~li 2 lg0!2y2g02# 1 aph0~lg!~li 2 lg!2~ g023! future moderate-resolution imaging spectroradiom-
3 ~ g 2 g0!exp@2~li 2 lg0!2y2g02# eter. Specifically, the four-band inversions em-
ployed bands located at, or near, 412, 490, 520, and
1 ad~ld!exp@2S0~li 2 ld!# 2 ad0~ld!exp@2S0~li 2 ld!# 565 nm. ~The 443-nm band was not used in the 4 3
3 ~li 2 ld!~S 2 S0! 1 bbt~lb!~lbyli !n0v~li ! 4 inversions because numerous unpublished analyses
of our airborne active–passive data suggest that the
1 ~n 2 n0!bbt0~lb!~lbyli !n0 ln~lbyli !v~li ! 443-nm band does not provide the best retrieval re-
5 2aw~li ! 2 bbw~li !v~li !. (6) sults when applied to field data. This finding is in
agreement with the research of others.3! The 5 3 5
Equation ~6! is linear in aph~lg!, ad~ld!, bbt~lb!, ~n 2 inversions utilized 412, 460, 490, 520, and 565 nm.
n0!, ~S 2 S0!, and ~ g 2 g0!. The oceanic-state vector The matrix inversion methodology can be adapted
of unknown IOP’s and parameters is thus p 5 easily to other band sets.
@aph~lg!, ad~ld!, bbt~lb!, ~n 2 n0!, ~S 2 S0!, ~ g 2 g0!#T. The inversion of the D matrix can be performed in
The initial or starting values for aph0, ad0, bbt0 can be several ways but we frequently use lower and upper
obtained by a standard 3 3 3 radiance model inver- triangular decomposition. Although we did not en-
sion1 using the g0, S0, and n0. The solution for p is counter any singularities, singular value decomposi-
given by the solution1 of Dp 5 h or p 5 D21 h. Of tion methods also provide a quantitative evaluation
course, the values of n, S, and g are immediately of any suspected singularities.4 The initial or start-
found by respectively adding the initial assumed val- ing value for each spectral model parameter was var-

1658 APPLIED OPTICS y Vol. 38, No. 9 y 20 March 1999

ied from 2100% to 100% with the exception that the
minimum initial guess for the phytoplankton spectral
model parameter g was 35 nm to avoid floating point
overflows in the computation of the exponential por-
tion of that spectral model. To simplify the process
of solving for a model parameter, the remaining IOP
spectral model parameters were held constant at
their fixed values ~ g 5 85 nm, S 5 0.017 nm21, or n 5
3.5! prior to each inversion.
After each pixel or radiance spectrum was in-
verted, the initial guess or starting value for the
parameter under study was reset to the original
value ~corresponding to a chosen percent offset from
the actual parameter!. This was done to fully test
how the technique responds to the wide range of
IOP’s used to create the synthetic radiances. If the
iterative inversion technique were to be used on Fig. 1. For the exponent n of the TCB coefficient wavelength ratio
actual data, the initial guess for the parameter be- spectral model: the percentage of the total data set ~;10,000
ing solved for would not be reset after each pixel. radiance spectra! that accurately converged when the initial or
Instead the retrieved value for that parameter from starting value n0 was a given percentage D away from the actual n
the previous pixel would be used for the new initial value used to generate the radiance spectrum. The starting value
value. For simulations, if the parameter value for n0 must be within a few percent of the actual value of n to
from the prior pixel was used as the initial value, it converge to the exact n value. These data suggest that lineariza-
would considerably limit the dynamic range of the tion and iterative matrix inversion not be used for retrieval of the
exponent n of the wavelength ratio backscatter model.
tests.
During the calculations, convergence was consid-
ered accomplished when the retrieved n 2 n0 ~or S 2 D. Retrieval of the Chromphoric Dissolved Organic
S0 or g 2 g0! changed by ,0.001%. An accurate Matter Spectral Slope S
retrieval was considered accomplished when the re-
trieved IOP spectral parameter ~n, S, or g! was within This inversion example retrieves S together with aph,
1% of the value used to generate the radiance spec- ad, and bbt in a 4 3 4 inversion. The CDOM spectral
slope parameter S was retrieved correctly for a fairly
trum. During execution, a self-imposed limit of 500
wide range of initial or starting values of S0. The
iterations is set, but the actual number seldom ex-
number of iterations needed for correct convergence
ceeded 50 iterations per pixel.
varied from ;3 to 20. Figure 2 shows the percent-
age of the total data set that accurately converged
C. Retrieval of the Total Constituent Backscattering
when the initial or starting value S0 was a given
Wavelength Ratio Exponent n
percentage D away from the S value used to generate
the radiance spectrum. If the starting value for S0
This first inversion example retrieves n together with was .225% from the actual value of S, then almost
aph, ad, and bbt in a 4 3 4 inversion. For initial or all iterative inversions successfully converged to the
starting values of n0 ~offset 2100% to 1100% from correct S value. When the initial value for S0 was
the actual value of n used to generate the radiances!, between 2100% and 225% from the actual S, then
100% correct convergence was infrequent. The the percent of correct convergence deteriorated rap-
number of iterations required to converge to within idly and most iterations did not converge to the cor-
1% of the exact n value ranged from 3 to 250. Figure rect S. These results suggest that using initial S0
1 shows the percentage of the total data set ~;10,000 values that are larger than the expected S provides a
radiance spectra! that accurately converged when the greater probability of correct convergence than initi-
initial or starting value n0 was a given percentage D ating S0 below the expected S. These data further
away from the actual n value. It can be seen that a suggest that the linearization procedure may be used
high percentage of correct convergence was obtained for retrieval of S provided that the starting value of
only when the initial guess n0 was near the actual S0 is generally chosen in the high range of expected S
value. Thus one must accurately know ~or guess! values.
the value of n before it can be satisfactorily retrieved.
If one knows ~or guesses! the value of n within a few E. Retrieval of the Phytoplankton Gaussian Width g
percent, there is little reason to perform the retrieval. This inversion example retrieves g together with aph,
Or, conversely, if one does not know the value of n, ad, and bbt in a 4 3 4 inversion. Retrieval of the
then it is highly likely that any guess will result in an phytoplankton spectral model parameter g somewhat
incorrect convergence. Accordingly, these data sug- mirrored the CDOM spectral model parameter S.
gest that the linearization procedure not be used for Figure 3 shows the percentage of the total data set
retrieval of the exponent n of the wavelength ratio that converged when the initial or starting value g0
backscattering model. was a given percentage D away from the actual g

20 March 1999 y Vol. 38, No. 9 y APPLIED OPTICS 1659

 continually declined. These data suggest that the lin-
earization procedure is more likely to retrieve a correct
spectral width g when the starting value g0 is chosen in
the low range of expected values of g ~but not less than
g 5 35 nm!.
F. Concurrent Retrieval of the Phytoplankton Gaussian
Width g and the Chromophoric Dissolved Organic Matter
Spectral Slope S
The previous 4 3 4 inversion results suggest that one
can simultaneously retrieve both g and S ~but not n!
together with the aph, ad, and bbt IOP’s in a 5 3 5 D
matrix solution similar to Eq. ~6!. ~Simultaneous
retrieval of spectral model pairs such as n, S and n, g
was not addressed because the retrieval of n, aph, ad,
and bbt was found to be largely unacceptable.! The
simulations showed that concurrent retrieval of S
Fig. 2. For the spectral slope S of the CDOM exponential spectral
and g ~together with the aph, ad, and bbt! was accept-
model: the percentage of the total data set ~;10,000 radiance
spectra! that accurately converged when the initial or starting
able when the guesses or initial values of S0 and g0
value S0 was a given percentage D away from the actual S value fall within certain ranges. Figure 4 provides a con-
used to generate the radiance spectrum. If the starting value for tour plot of the correct convergence percentage and
S0 was .225% from the actual value of S, then almost all iterative the ranges of S0 and g0 relative to their actual values.
inversions successfully converged to within 1% of the S value used One must choose pairs of S0 and g0 so that they fall
to generate the radiance spectrum. These data suggest that the comfortably within the 95–100% contour regions in
linearization procedure is likely to retrieve a correct S when the Fig. 4.
starting value S0 is chosen in the high range of expected S values.
4. Discussion
It is generally agreed that IOP’s of multiple oceanic
value. It can be seen that virtually all the inversions constituents cannot be retrieved by empirical radiance
converged to within 1% of the correct g if the initial ratios. Worse yet, IOP spectral model parameters are
guess for g0 was within 265% to 155%. The number virtually inaccessible by empirical radiance ratio anal-
of iterations needed for this range of g varied from 3 to yses. To mitigate the significant deficiencies of sim-
11. As the offset of g0 from g increased past 55%, the ple band ratios, a radiance model was recast in matrix
number of inversions that converged to the correct g form by exploiting the linear summability of the IOP’s.
Then the linear matrix inversion theory, normally
used to retrieve the aph, ad, and bbt IOP’s, was ex-
tended to include spectral model parameters such as
the TCB coefficient wavelength ratio model exponent
n, the CDOM absorption coefficient spectral slope S, or
the phytoplankton absorption Gaussian model spec-
tral width g individually. ~It was then shown that, in
principle, any number of spectral model parameters
and their associated IOP’s can be accommodated
within the linear matrix retrieval formalism.! The
individual retrieval of n or S or g ~concurrently with
aph, ad, and bbt! was then studied in detail by using 4 3
4 matrix inversion.
The exponent of the TCB wavelength ratio spectral
model required a narrow range of acceptable starting
values to achieve convergence to the correct value for
n. Specifically, one must provide a starting value for
n0 that is within a few percent of the actual value for
n; otherwise convergence to the correct exponent
Fig. 3. For the spectral width g of the phytoplankton Gaussian value cannot be assured even with error-free radi-
spectral model: the percentage of the total data set ~;10,000 ance data. Because recent research suggests that n
radiance spectra! that accurately converged when the initial or can vary from 0 –2.4 in some waters5 and 2.0 – 4.1 in
starting value g0 was a given percentage D away from the actual g
other waters,6 it is unlikely that one can initiate the
value. If the starting value for g0 was within 255% and 55% of
the actual value of g, then all iterative inversions successfully retrieval of n with a sufficiently accurate n0 to im-
converged to within 1% of the g value used to generate the radiance prove the knowledge of n. Accordingly, the retrieval
spectrum. These data suggest that the linearization procedure is of n is not recommended because it is unlikely that
likely to retrieve a correct spectral width g when the starting value one can initiate the iterative calculation with suffi-
g0 is chosen in the low range of expected values of g. cient accuracy to merit the attempt. Thus one must

1660 APPLIED OPTICS y Vol. 38, No. 9 y 20 March 1999

Fig. 4. For width g of the phytoplankton Gaussian spectral model and slope S of the CDOM exponential spectral model: the percentage
of the total data set ~;10,000 radiance spectra! that accurately converged when the initial or starting values g0 and S0 were a given
percentage D away from the actual g and S values used to generate the radiance data. Correct convergence is found within the 95–100%
contour region. These data suggest that the linearization procedure is more likely to retrieve both a correct spectral width g and a CDOM
slope S when starting values g0 and S0 are chosen to be larger than their expected values.

either ~a! choose a fixed value for n or ~b! develop a get rest heavily on improved knowledge of the TCB
separate algorithm to retrieve n. If these latter two spectral model.
approaches do not provide satisfactory IOP retriev- The simulations suggested that retrievals of the
als, then an entirely new model for the TCB coeffi- CDOM spectral slope S ~together with the TCB,
cient must be developed. The simulations herein CDOM, and phytoplankton IOP’s! have a higher
vividly show that the TCB coefficient model is a gen- probability of accurate convergence if the initial or
uine source of concern. But it is clear that it will be starting values of S0 are chosen to be larger than the
an even greater concern when applied to real data. expected value of S. Conversely, the simulations
In this latter situation, the TCB model predicament suggested that retrievals of phytoplankton Gaussian
is exacerbated by the fact that all the constituent width g ~together with the TCB, CDOM, and phyto-
scatterers are grouped into one spectral model. As plankton IOP’s! have a higher probability of accurate
the concentrations of the constituent scatterers convergence if the initial or starting values of g0 are
change relative to each other, the TCB spectral model chosen to be smaller than the expected value of g.
must, by necessity, vary its slope to accurately rep- Unlike the TCB spectral parameter n, the CDOM
resent the varying combinations of the scatterers. spectral parameter S or the phytoplankton spectral
Until separate backscattering models are developed model g are recommended for individual retrievals
for each scattering constituent, the TCB spectral using the linearization procedure and subsequent 4 3
model approach will remain a constant annoyance in 4 inversion.
radiance model inversions. For this reason, it is With the retrieval of n excluded from consideration,
suggested that further efforts be expended to improve the retrieval of the g and S spectral model parameters
the retrieval of n ~andyor the entire TCB spectrum!. ~together with three IOP’s! was found to be possible.
Another reason for focusing on the TCB model is that Specifically, using five sensor bands to simultaneously
uncertainty in the exponent n strongly influences the solve for two spectral parameters and three principal
CDOM absorption coefficient retrieval uncertainty.1 IOP’s, the 5 3 5 inversions were successful when the
Thus better estimates of the global ocean carbon bud- starting values S0 and g0 were chosen carefully.

20 March 1999 y Vol. 38, No. 9 y APPLIED OPTICS 1661

 Future studies are recommended to address the 2. C. R. Wylie, Jr., Advanced Engineering Mathematics, 3rd ed.
quantification of errors in retrieved spectral model ~McGraw-Hill, New York, 1960!, Chap. 4, pp. 136 –137.
parameters caused by radiance measurement errors 3. J. Aiken, G. F. Moore, C. C. Trees, S. B. Hooker, and D. K. Clark,
“The SeaWiFS CZCS-type pigment algorithm,” NASA Tech.
and other spectral model parameter uncertainties.
Memo. 104566, S. B. Hooker and E. R. Firestone, eds. ~NASA
Such analyses are necessary to further establish that Goddard Space Flight Center, Greenbelt, Md., 1995!, Vol. 29,
the matrix methodology is practical and that unde- p. 1.
sirable properties, if any, are identified. The study 4. W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flan-
and potential use of optimum iterative methods is nery, Numerical Recipes in Fortran, 2nd ed. ~Cambridge U.
also suggested for further study of this methodology. Press, New York, 1992!, Chap. 15, p. 672.
Details of these iterative methods can be found in 5. Z. P. Lee, K. L. Carder, S. K. Hawes, R. G. Steward, T. G.
numerical analysis references.4 Peacock, and C. O. Davis, “Model for the interpretation of hy-
perspectral remote-sensing reflectance,” Appl. Opt. 33, 5721–
References 5732 ~1994!.
1. F. E. Hoge and P. E. Lyon, “Satellite retrieval of inherent optical 6. R. A. Maffione, D. R. Dana, and J. M. Voss, “Spectral depen-
properties by linear matrix inversion of oceanic radiance mod- dence of optical backscattering in the ocean,” presented at OSA
els: an analysis of model and radiance measurement errors,” Annual Meeting, Portland, Oregon, 10 –15 September 1995, pa-
J. Geophys. Res. 101, 16,631–16,648 ~1996!. per MDD4, p. 56.

1662 APPLIED OPTICS y Vol. 38, No. 9 y 20 March 1999