## Abstract

This work presents a novel GA-SA approach to retrieve the constituents of water
bodies from remote sensing of ocean color. This approach is validated and
compared to the existing algorithms using the same synthetic and
*in-situ* datasets compiled by the International Ocean Color
Coordinate Group. Comparing to the other methods, the GA-SA approach provides
better retrievals for both the inherent optical properties and various water
constituents. This novel approach is successfully applied in processing the
images taken by MODerate resolution Imaging Spectroradiometer (MODIS) and
generates regional maps of chlorophyll-*a* concentration, total
suspended matter, and the absorption coefficient of color dissolved organic
matter at 443nm.

©2007 Optical Society of America

## 1. Introduction

Retrieving information of water constituents is an extremely important application of
ocean color analysis. For instance, retrieved chlorophyll concentration is essential
for estimating the primary productivity of an ocean [1]. Over the past two decades, most studies relied on
regression analysis to derive the empirical relationship between chlorophyll
(*Chl*) concentration and the above-surface remote sensing
reflectance *R _{rs}*(

*λ*). For application on a global scale, however, the applicability of regression algorithms is limited to the local region for which empirical relationships are derived. Semianalytical (SA) models [2] can simulate sea surface reflectance (SSR) based on the fundamental physical principles and can account for measurement uncertainties [3]. Therefore, numerous retrieval algorithms shifted toward coupling SA models to simulate apparent optical properties (AOP), and then inversely retrieving the constituents of water bodies via optimization techniques [3, 4] such as the Levenberg-Marquart approach [5], and nonlinear least-square schemes [6, 7]. Alternatively, SA paradigms can also be modified to a simplified expression for linear inversion, such as the simplex method [8] and the matrix inversion based on the least-square approach [9].

Consensus has been reached in recent years that retrieval of inherent optical
properties (IOP) is theoretically straightforward and more accurate than retrieving
water constituents on a global scale [10]. For practical applications that require not only IOPs but
also the constituents of water bodies, however, it remains necessary to further
decompose contributions to total IOPs from individual water constituents based on
biooptical models (BOM). The BOMs are comprised of various formulations and
parameters that are generally nonlinear and vary globally [3]. Retrieving water constituents from complex BOMs is a
challenging task as conventional optimization approaches might reduce their accuracy
without constraining initial values [11] or limiting the number of decision variables [6]. Consequently, soft computing and global optimization
methods, such as simulated annealing [3], neural networks [12] and genetic algorithms [11] (GA), were introduced in numerous works to resolve this
inverse problem. The advantage of GA over conventional optimization methods were
demonstrated by Zhan *et al*. [11]. For example, GA is able to retrieve IOPs with limited
amount of *a priori* information. It is particularly suitable for
solving problems of complex model for which the interaction of parameters is highly
non-linear [13].

This work presents a novel GA-SA approach for retrieving water constituents from
remote sensing of ocean color under the assumption that other IOPs covary with
phytoplankton absorption coefficient at 440nm *a _{ph}*(440).
Application of this novel GA-SA approach is validated against a synthetic dataset
(N=500) and an

*in-situ*dataset (N=656) compiled by the International Ocean Color Coordinate Group (IOCCG) [14]. Note that the synthetic dataset comprises a wide range of parameters characterizing the global ocean (Case 1 waters), and most of the

*in-situ*data come from locations that are relatively close to the coast (some are Case 2 waters) [15]. Nevertheless, the GA-SA approach can retrieve

*a*(440) (linear percentage error, 40%) for synthetic data and

_{ph}*a*(443) (linear percentage error, 86%) for

_{ph}*in-situ*data. Compared with existing retrieval algorithms published in the recent IOCCG report [14], the proposed GA-SA approach provides better retrievals of water constituents and IOPs, and can be implemented to process images taken by MODerate resolution Imaging Spectroradiometer (MODIS) on a regional scale. This work encourages application of the proposed GA-SA approach for deriving other products of ocean color at a regional scale, such as the color dissolved organic matter (CDOM) and non-algal particle/detritus/mineral (NAP).

## 2. Semianalytical model

Lee *et al*. [16] presents a full description of the SA model. Briefly, the
SA model simplifies the radiative transfer process by relating
*R _{rs}*(

*λ*) to the dimensionless number

*u*(

^{rs}*λ*) as

where *r _{rs}*(

*λ*) is the below-surface remote-sensing reflectance spectra;

*g*

_{0}=0.0895 and

*g*

_{1}=0.1247 are taken from Lee

*et al*. [17]. Note that

*u*(

*λ*) derived from

*r*(

_{rs}*λ*) is denoted as

*u*(

^{rs}*λ*). In addition,

*u*(

*λ*) can also be derived by a function of the absorption coefficient

*a*(

*λ*) and the backscattering coefficient

*bb*(

*λ*) [18], both of which can be derived from in-water constituents

The distinction of *u ^{rs}*(

*λ*) and

*u*(

^{iwc}*λ*) is just based on how they are derived, for the purpose of forming an explicit objective function of optimization in the subsequent section.

Although the retrievals of *a*(*λ*) and
*bb*(*λ*) from
*R _{rs}*(

*λ*) are robust and accurate, to further decompose the contributions to total

*a*(

*λ*) and

*bb*(

*λ*) from individual water constituents still requires a set of BOMs that are sufficiently flexible to characterize the wide variety of biooptical relationships among various water constituents [19]. The four-component BOM [20] utilized in this research is summarized as follows. The BOM is typically a function of

*Chl*[21] or

*a*(

_{ph}*λ*). This research follows the work by Lee

*et al*. [16] to parameterize

*a*(

_{ph}*λ*) as

where the empirical coefficients for
*a _{0}*(

*λ*) and

*a*(

_{1}*λ*) are in Table 2 of Lee

*et al*. [16]. The optical properties of additional water constituents are generally also co-varied with

*a*[21]. Therefore, this work uses the similar co-variation formulations for an ocean color model for global applications [3] to yield a general expression of optical properties contributed of other constituents

_{ph}In this formulation, *R _{i}* is the covariance between

*a*(440) and other IOPs (

_{ph}*X*). Thus, there would be four combinations of

_{i}*X*and

_{i}*R*, i.e.,

_{i}*R*is the ratio between the absorption coefficient of CDOM at 440nm

_{g}*a*(440) and

_{g}*a*(440);

_{ph}*R*is the ratio between the absorption coefficient of NAP at 440nm

_{d}*a*(440) and

_{d}*a*(440);

_{ph}*R*is the ratio between the backscattering coefficient of phytoplankton at 550nm

_{bph}*bb*(550) and

_{ph}*a*(440), and

_{ph}*R*is the ratio between the backscattering coefficient of NAP at 550nm

_{bd}*bb*(550) and

_{d}*a*(440). Except for the contribution of

_{ph}*a*(

_{ph}*λ*) and the absorption of pure water [22]

*a*(

_{w}*λ*), the absorption coefficients for other water constituents can be modeled as an exponential function of wavelength [23]. Thus, five parameters are used to express

*a*(

*λ*) as

$$\phantom{\rule{1.6em}{0ex}}+{a}_{g}\left(440\right)\mathrm{exp}\left[-{S}_{g}\left(\lambda -440\right)\right]+{a}_{d}\left(440\right)\mathrm{exp}\left[-{S}_{d}\left(\lambda -440\right)\right],$$

where *S _{g}* and

*S*are the spectral slope of absorption coefficients for CDOM and NAP, respectively. To compare with the IOP derived by other algorithms in subsequent sections, the absorption coefficient due to NAP and CDOM

_{d}*a*(

_{dg}*λ*) is calculated by the summation of

*a*(

_{g}*λ*) and

*a*(

_{d}*λ*). Similarly, the backscattering coefficients of other water constituents can be modeled as a power law function [16]. Therefore, four parameters are used to express

*bb*(

*λ*) as

where *Y _{ph}* and

*Y*are the spectral shapes of backscattering coefficients for phytoplankton and NAP, respectively, and

_{d}*b*(

_{w}*λ*) is the scattering coefficient of pure seawater [24]. To compare with the IOP derived by other algorithms in subsequent sections, the backscattering coefficient due to particles

*bb*(

_{p}*λ*) is calculated by the summation of

*bb*(

_{ph}*λ*) and

*bb*(

_{d}*λ*).

## 3. GA-SA approach

The theoretical foundations of GAs and details of the searching approach for optimization are found in Goldberg [13]. Briefly, the GAs are stochastic algorithms with searching approaches that mimic phenomena in nature, such as inheritance and Darwinian struggle for survival. To solve inverse problems, the concept of GA can be utilized to optimize a quantifiable objective, namely, fitness. The fitness value is an evaluation of the solutions purposed for the problem. Solutions with high fitness values would have high probabilities to survive in the revolution process. Evaluation of fitness depends on “decision variables” of the forward model, which are usually nonlinear and must subject to “constraints.”

Therefore, this work has three key points in combining GAs and SA to solve the
inverse problem in ocean optics. Because
*u ^{rs}*(

*λ*) can be obtained semi-analytically from SSR with few errors [17] [Eqs. (1) and (2)], the first key point is to calculate

*u*(

^{iwc}*λ*) in Eq. (3). Nine unknown variables in Eq. (4) to Eq. (7) need to be decided before Eq. (3) can be calculated, which are so-called decision variables, i.e.,

*a*(440),

_{ph}*R*,

_{g}*R*,

_{d}*S*,

_{g}*S*,

_{d}*R*,

_{bph}*R*,

_{bd}*Y*, and

_{ph}*Y*. Since the minimal divergence between

_{d}*u*(

^{iwc}*λ*) and

*u*(

^{rs}*λ*) represents a good agreement between the actual amount of water constituents and the retrievals from remote sensing. Therefore, the second point is to minimize

*obj*and with an intention to obtain high fitness values, as

where *N* is the number of available satellite bands and
*i* is the band index. GA searches the best set of decision variables
that falls within the range specified by users, which means that GA would yield a
near-best solution if the best set of decision variables does not exist in the
searchable range. For example, GA would generate a solution with
*a _{ph}*(440) equals to 0.5 if the actual solution of

*a*(440) is 0.6 and the upper bound for searching is 0.5. Additionally, extra computing time is required for GA to find the best solution in a broader range. To achieve a better performance and efficiency, the third point is to constrain all parameters in a dynamic range when

_{ph}*u*(

^{iwc}*λ*) is examined. In this work, the constraints on the range of all decision variables are determined by Lee’s model [15] that was specifically designed to generate a comprehensive synthetic data set based on observation and theory (Table 1). Note that some parameters are specified in GA to control the efficiency of GA searching (see Table 2). The initial ranges of these parameter values are referred to practical rules [13], and their final ranges are determined by trial and error tests. The more accurate solution can be obtained with large number of generations at the cost of more computing time. Fig. 1 presents a flowchart and some examples of the GA operators to illustrate the procedures of GA-SA; these procedures are summarized as follows.

- 1. Each GA run comprises a number of generations in a population of individuals. Each individual, which represents one possible solution to the SA inverse problem, can be any combinations of parameters to be determined.
- 2. Since GA attempts to identify the best individual in a genetic manner, the GA-SA process begins by randomly generating a population of individuals (chromosomes, strings). In the decoding example of Fig. 1, three alphabets (0 or 1) are used to represent one real value of each decision variable, therefore, there are 8 (2
^{3}) possible values for each of them. The chromosome is a combination of nine decision variables, which comprises 27 alphabets (genes) to represent one possible solution of the inverse problem. A binary decoding skill is utilized to translate an individual from a chromosome form to a real form, as$${x}_{i}={x}_{i}^{min}+\frac{{x}_{i}^{max}-{x}_{i}^{min}}{{2}^{{l}_{i}}-1}\times \sum _{j=1}^{l}\left({2}^{j-1}\xb7{\mathit{bit}}_{\mathit{ij}}\right),$$where

*x*is the real value of the_{i}*i*th parameter;*x*is the upper bound of the_{i}^{max}*i*th parameter;*x*is the lower bound of the_{i}^{min}*i*th parameter;*l*is the string length of the_{i}*i*th parameter;*j*is the index number of genes;*bit*is the alphabet in binary coding (0 or 1) of the_{ij}*j*th gene of the*i*th parameter. The fitness value is then assessed by the “real form” of the individual using BOMs, a given*u*(^{rs}*λ*), and Eq. (8). - 3. The second and subsequent generations are created via a reproduction process. In this process, an individual with a high fitness value has a likely opportunity to be selected and pass its genetic characteristics to its offspring by duplicating or processing a “crossover operation” with other selected individuals (Fig. 1). The crossover operator is applied to create offspring by randomly selecting the corresponding genes of two parent chromosomes for exchanging genetic features. Following the crossover operation, offspring have various genetic features from their parents and in some cases, “good” features are destroyed, resulting in an ineffective search. A GA parameter called crossover probability (Pc) is introduced to obtain a balance between diversity and search efficiency. When Pc=0.5, approximately 50% of selected individuals in the previous generation participate the crossover process and the remaining individuals pass on their genetic features to the next generation.
- 4. Retaining the diversity of genetic features is the most important goal that prevents the process of numerical evolution ending at a local solution. To achieve this aim, a “mutation operator” is utilized for the reproduction process by randomly varying genetic features (Fig. 1). The mutation operator dramatically changes the genetic features; thus, the probability of mutation is commonly considerably lower than Pc, as is the case for humans.
- 5. In the final generation, the optimum set of all decision variables for given
*u*(^{rs}*λ*) can be obtained. With this optimum set, all required information for IOPs can be derived and the inverse problem is solved.

## 4. Results and discussion

To evaluate and compare the performance of various retrieval algorithms, the IOCCG
compiled a synthetic dataset [25] and an *in-situ* dataset containing IOPs and
corresponding AOPs [26]. Uncertainties in IOP measurements were integrated into the
synthetic dataset by mapping one *Chl* to 25 different sets of IOPs
based on theory and field data [15], resulting in 500 IOPs that are generated from 20 levels of
*Chl* in the range of 0.03–30 mg/m^{3}, and
the corresponding SSRs simulated by Hydrolight for the sun at 30° from
zenith. The *in-situ* dataset comprises 656 cases originating from
NASA’s SeaWiFS Bio-optical Archive and Storage System (SeaBASS),
including Chl-*a*, *a*(*λ*),
*a _{dg}*(

*λ*),

*a*(

_{ph}*λ*) and

*R*(

_{rs}*λ*) at the first five SeaWiFS bands (

*λ*=412, 443, 490, 510 and 555nm). Utilizing these two datasets, Lee

_{i}*et al*. [14] assessed and compared the performance of nine retrieval algorithms by calculating the Root-Mean-Square-Error in log phase (

*RMSE*). The

*RMSE*value for these retrieval algorithms were taken to calculate the linear percentage errors ε in log scale [17], as

Tables 3 and 4 present summaries of the *ε* for
retrieving *a*(*λ*),
*a _{dg}*(

*λ*),

*bb*(

_{p}*λ*) and

*a*(

_{ph}*λ*). Note that

*bb*is not included in the

_{p}*in-situ*data. This work employs the same datasets and applies the same procedures as described by

*Lee et al*. [14] to validate the proposed GA-SA model.

#### 4.1 Test using the synthetic dataset

Figure 2 and Table 3 present the performance of the proposed GA-SA by
comparing the retrievals of 500 synthetic cases to the known IOPs. For the
purpose of deriving water constituents, a distinction between individual and
combinational IOPs is required. The combinational IOP contains the optical
information attributed to more than one water constituent, e.g.,
*a*, *bb*, *a _{dg}*, and

*bb*. The individual IOP, e.g.,

_{p}*a*and

_{ph}*bb*, is needed for accurately deriving the concentration of the specified water constituent. For the category of combinational IOPs, most of the existing algorithms are able to attain a good retrieval. Notably, the Quasi Analytical Algorithm [17] (QAA) obtains the best retrievals for

_{d}*a*in this category, whereas the proposed GA-SA approach ranks second among all existing algorithms and provides the best retrieval of

_{dg}*bb*(550). As illustrated in Table 3,

_{p}*a*(

_{ph}*λ*) retrievals is approximately 1.6–6.7 times worse than the

*a*(

*λ*) retrievals when using other algorithms. The results of Figs. 2(a) and 2(e)-2(g) also show that it is difficult for GA-SA to divide the contributions between

*a*(

_{ph}*λ*),

*a*(

_{g}*λ*) and

*a*(

_{d}*λ*), from

*a*(

*λ*) to obtain the retrievals of individual IOPs. However, the proposed GA-SA still yields a good result for

*a*(440) with

_{ph}*ε*as low as 40%, which is only 1.5 times the error rate for

*a*(

*λ*) retrievals. This result demonstrates that the GA-SA approach is ideal for further decomposing the contributions of individual water constituents to total

*a*(

*λ*) and

*bb*(

*λ*).

Another significant advantage of the proposed GA-SA approach is that
*a _{g}*(440),

*a*(440),

_{d}*bb*(550) and bbd(550) can be retrieved simultaneously. As the spectral shape of one IOP might be very similar to the spectral shape of the other IOPs, e.g.,

_{ph}*a*and

_{d}*a*in Eq. (5) ,and

_{g}*bb*and

_{ph}*bb*in Eq. (6), a relatively high contribution comes from one IOP would overlap the contribution from the other IOPs. Therefore, the overlapping absorption and backscattering makes it difficult to differentiate individual IOPs from the total optical properties using the general SA models. In the synthetic dataset, for example, this is particularly true for

_{d}*a*(440) that accounts for 12% of

_{d}*a*(440), and

*bb*(550) that accounts for 35% of

_{ph}*bb*(550). By contrast,

*a*(440) accounts for 64% of

_{g}*a*(440), and

*bb*(550) accounts for 61% of

_{d}*b*(550). As a result, the values of R

_{b}^{2}for retrieving

*a*(440) and

_{g}*bb*(550) are as high as 0.965 and 0.870, respectively. Despite of the relative low contribution of

_{d}*a*(440) (23%) to

_{ph}*a*(440), the GA-SA approach is able to retrieve this key component with the lowest error among all IOPs. This is because

*a*(440) has a particular BOM [Eq. (4)] and most variables are directly or indirectly associated with

_{ph}*a*(440).

_{ph}#### 4.2 Test using the in-situ dataset

Figure 3 and Table 4 show the performance of the GA-SA approach by
comparing the retrievals of GA-SA approach to the 656 *in-situ*
cases of measured IOPs. Due to the high uncertainty in measurement errors in the
*in-situ* dataset, the retrieving errors (Table 4) of the QAA is increased from 2.4 (for
*a _{ph}*) to 3.3 (for

*a*) times to the results of the 500 synthetic cases (Table 3). Conversely, the GA-SA approach is relatively robust with only 2.0 (for

*a*) to 2.2 (for

_{dg}*a*) times to the results of the 500 synthetic cases. To retrieve the combinational IOPs, the GA-SA approach is ranked second and is only slightly less accurate than the QAA. In retrieving the key component

*a*(443), however, the GA-SA scheme performs best. Compared with algorithms that disregard several cases due to an error checking mechanism, GA-SA performance was evaluated for all 1156 cases—synthetic and

_{ph}*in-situ*datasets combined. No unreasonable retrieval (i.e., negative IOPs) was found in any case.

#### 4.3 Applications to processing satellite imagery

The proposed GA-SA approach is accurate and robust, and easy to use and sufficiently fast at processing satellite imagery on a regional scale. Figure 4 presents an example of Taiwan’s coastal region taken by MODIS-Aqua on May 9, 2006. Although the GA-SA approach is applied in a pixel-by-pixel manner, the processed image is fairly smooth without apparent discontinuity. All processing is conducted on a personal computer equipped with a Pentium 2.4GHz CPU. It took 3.6 hours to process 15,000 pixels. Processing can be accelerated by parallelizing the computation on a cluster machine.

Compared to the standard product of MODIS, the proposed GA-SA approach yields
comparable values of *Chl* [Figs. 4(a) and 4(b)], except for coastal waters where the NAP or CDOM
dominates the spectral shape of SSR [29]. This is exactly the problem in existing global
algorithms that typically overestimate *Chl* in coastal waters.
In order to clarify the contributions to SSR from *Chl*, NAP and
CDOM, the IOPs of individual water constituents need to be retrieved as accurate
as possible. As shown in Fig. 2, the retrievals of
*a _{ph}*(440) and

*bb*(550) derived from the GA-SA approach are more representative and accurate, comparing to other retrievals, such as

_{d}*bb*(550) and

_{ph}*a*(440). Therefore, the

_{d}*Chl*(mg/m

^{3}) is calculated by [30]

and the NAP (g/m^{3}) is calculated by [31]

where *B _{d}* depends on the selected phase function, and
is 0.0183 when Petzold average particle phase function [32] is used. Finally, the concentration of CDOM
(m

^{−1}) is expressed by

*a*(443) that is calculated by the retrievals of

_{g}*a*(440),

_{ph}*R*and Eq. (6).

_{g}One remarkable feature shown in Fig. 4 is the river plumes along Taiwan’s west
coast [referring to the two red frames in Fig. 4(a)]. The standard product of *Chl* [Fig. 4(a)] does show the distribution and dispersion of
the plumes; however, the high values of *Chl* seem to be not
realistic and no information for NAP and CDOM exists. Conversely, the coastal
GASA value [Fig. 4(b)] is roughly of the same order as the retrieval
of *Chl* in offshore waters, except for two major estuaries for
which the concentrations of *Chl* seem to be generally high
[referring to the two arrows in Fig. 4(b)]. Furthermore, the maps for NAP [Fig. 4(c)] and CDOM [Fig. 4(d)] derived using the GA-SA approach clearly
delineate the distributions and dispersions of several major river plumes. These
two new products facilitate the study of river plume dynamics and the estimation
of the annual amount of sediment discharged through the river-sea system.

## 5. Summary

This work presents a novel GA-SA approach for retrieving water constituents from
remote sensing of ocean color. This approach is validated and evaluated against a
synthetic data set and an *in-situ* data set compiled by IOCCG [14]. The result of comparison demonstrates that the GA-SA
approach is as almost as accurate as the QAA method in retrieving combinational
IOPs. Furthermore, the proposed approach is particularly ideal for further
decomposing the contributions from individual water constituents to total IOPs. The
GA-SA approach is accurate and robust, and is easily utilized and sufficiently fast
at processing the satellite imagery on a regional scale. After processing by the
GA-SA approach, the MODIS-Aqua image of Taiwan’s coastal region
(2006/5/9) yields a realistic map of *Chl*, both for coastal and
offshore waters. Furthermore, the maps of NAP and CDOM clearly delineate the
distributions and dispersions of several major river plumes. This work suggests that
the constituents of water bodies, such as *Chl*, NAP and CDOM, can be
acquired routinely from remote sensing of ocean color.

## 6. Acknowledgment

The authors would like to thank the National Science Council of the Republic of China, Taiwan, for financially supporting this research under Contract Nos. NSC-95-2625-Z-006-004-MY3, NSC-95-2211-E006-283 and NSC-95-2611-M-006-002.

## References and links

**1. **T. Platt, “Primary production of the ocean
water column as a function of surface light intensity: algorithms for remote
sensing,” Deep-Sea Res. Part I-Oceanogr.
Res. Pap. **33**,149–163
(1986). [CrossRef]

**2. **H. R. Gordon, O. B. Brown, and M. M. Jacobs, “Computed relationships between the
inherent and apparent optical properties of a flat homogeneous
ocean,” Appl. Opt. **14**,417–427
(1975). [CrossRef] [PubMed]

**3. **S. Maritorena, D. A. Siegel, and A. R. Peterson, “Optimization of a semianalytical
ocean color model for global-scale applications,”
Appl. Opt. **41**,2705–2714
(2002). [CrossRef] [PubMed]

**4. **Z. P. Lee, K. L. Carder, C. D. Mobley, R. G. Steward, and J. S. Patch, “Hyperspectral remote sensing for
shallow waters: 2. Deriving bottom depths and water properties by
optimization,” Appl. Opt. **38**,3831–3843
(1999). [CrossRef]

**5. **K. Y. Kondratyev, D. V. Pozdnyakov, and L. H. Pettersson, “Water quality remote sensing in the
visible spectrum,” Int. J. Remote Sens. **19**,957–979
(1998). [CrossRef]

**6. **S. A. Garver and D. A. Siegel, “Inherent optical property inversion
of ocean color spectra and its biogeochemical interpretation .1. Time series
from the Sargasso Sea,” J. Geoph. Res. **102**,18607–18625
(1997). [CrossRef]

**7. **C. C. Liu and R. L. Miller, “A Spectrum Matching Method for
Estimating the Inherent Optical Properties from Remote Sensing of Ocean
Color,” in *Ocean Remote Sensing and
Imaging II*, R. J. Frouin, G. D. Gilbert, and D. Pan, eds. (San Diego, California
USA, 2003), pp.
141–152.

**8. **R. Doerffer and J. Fischer, “Concentrations of chlorophyll,
suspended matter, and gelbstoff in Case-II waters derived from Satellite
Coastal Zone Color Scanner data with inverse modeling
methods,” J. Geoph. Res. **99**,7457–7466
(1994). [CrossRef]

**9. **F. E. Hoge and P. E. Lyon, “Satellite retrieval of inherent
optical properties by linear matrix inversion of oceanic radiance models: An
analysis of model and radiance measurement
errors,” J. Geoph. Res. **101**,16631–16648
(1996). [CrossRef]

**10. **J. R. V. Zaneveld, “A theoretical derivation of the
dependence of the remotely-sensed reflectance of the ocean on the inherent
optical-properties,” J. Geoph. Res. **100**,13135–13142
(1995). [CrossRef]

**11. **H. G. Zhan, Z. P. Lee, P. Shi, C. Q. Chen, and K. L. Carder, “Retrieval of water optical
properties for optically deep waters using genetic
algorithms,” IEEE Trans. Geosci. Remote
Sensing **41**,1123–1128
(2003). [CrossRef]

**12. **A. Tanaka, M. Kishino, R. Doerffer, H. Schiller, T. Oishi, and T. Kubota, “Development of a neural network
algorithm for retrieving concentrations of chlorophyll, suspended matter and
yellow substance from radiance data of the ocean color and temperature
scanner,” J. Oceanogr. **60**,519–530
(2004). [CrossRef]

**13. **D. E. Goldberg, *Genetic Algorithms in Search, Optimization and Machine
Learning* (Addison-Wesley, New York,1989).

**14. ** IOCCG, “Remote Sensing of
Inherent Optical Properties: Fundamentals, Tests of Algorithms, and
Applications,” in *Reports of the
International Ocean-Colour Coordinating Group*, No. 5, Z. P. Lee, ed. (IOCCG, Dartmouth,2006).

**15. **Z. P. Lee, “Models, parameters, and approaches
that used to generate wide range of absorption and backscattering
spectra,” (IOCCG,2003),
http://www.ioccg.org/groups/lee_data.pdf.

**16. **Z. P. Lee, K. L. Carder, C. D. Mobley, R. G. Steward, and J. S. Patch, “Hyperspectral remote sensing for
shallow waters. I. A semianalytical model,”
Appl. Opt. **37**,6329–6338
(1998). [CrossRef]

**17. **Z. P. Lee, K. L. Carder, and R. A. Arnone, “Deriving inherent optical properties
from water color: a multiband quasi-analytical algorithm for optically deep
waters,” Appl. Opt. **41**,5755–5772
(2002). [CrossRef] [PubMed]

**18. **H. R. Gordon, O. B. Brown, R. H. Evans, J. W. Brown, R. C. Smith, K. S. Baker, and D. K. Clark, “A semianalytic radiance model of
ocean color,“ J. Geoph. Res. **93**,10909–10924
(1988). [CrossRef]

**19. **C. C. Liu, “Fast and accurate model of
underwater scalar irradiance for stratified Case 2
waters,“ Opt. Express **14**,1703–1719
(2006). [CrossRef] [PubMed]

**20. **R. P. Bukata, J. H. Jerome, K. Y. Kondratyev, and D. V. Pozdnyakov, *Optical Properties and Remote Sensing of Inland and
Coastal Waters* (CRC Press, Boca Raton,1995).

**21. **A. Morel, “Optical modeling of the upper ocean
in relation to its biogenous matter content (Case I
waters).“ J. Geoph. Res. **93**,10749–10768
(1988). [CrossRef]

**22. **R. Pope and E. Fry, “Absorption spectrum (380
– 700 nm) of pure waters: II. Integrating cavity
measurements,“ Appl. Opt. **36**,8710–8723
(1997). [CrossRef]

**23. **K. L. Carder, F. R. Chen, Z. P. Lee, S. K. Hawes, and D. Kamykowski, “Semianalytic MODerate-resolution
Imaging Spectrometer algorithms for chlorophyll a and absorption with
bio-optical domains based on nitrate-depletion
temperatures,“ J. Geoph. Res. **104**,5403–5421
(1999). [CrossRef]

**24. **A. Morel, ed. *Optical Properties of Pure Water and Pure Sea
Water* (Academic, New York,1974).

**25. **Z. P. Lee, “Ocean-colour
algorithms.“
http://www.ioccg.org/groups/OCAG_data.html.

**26. **Z. P. Lee, “Results from the IOCCG ocean color
algorithm working group.“
http://seabass.gsfc.nasa.gov/ioccg.html.

**27. **H. Schiller and R. Doerffer, “Improved determination of coastal
water constituent concentrations from MERIS
data,“ IEEE Trans. Geosci. Remote Sensing **43**,1585–1591
(2005). [CrossRef]

**28. **E. Devred, C. Fuentes-Yaco, S. Sathyendranath, C. Caverhill, H. Maass, V. Stuart, T. Platt, and G. White, “A semi-analytic seasonal algorithm
to retrieve chlorophyll-a concentration in the Northwest Atlantic Ocean from
SeaWiFS data,“ Indian J. Mar. Sci. **34**,356–367
(2005).

**29. **K. L. Carder, S. K. Hawes, K. A. Baker, R. C. Smith, R. G. Steward, and B. G. Mitchell, “Reflectance model for quantifying
chlorophyll-a in the presence of productivity degradation
products,“ J. Geoph. Res. **96**,20599–20611
(1991). [CrossRef]

**30. **A. Bricaud, M. Babin, A. Morel, and H. Claustre, “Variability in the
chlorophyll-specific absorption-coefficients of natural phytoplankton -
analysis and parameterization,“ J. Geoph.
Res. **100**,13321–13332
(1995). [CrossRef]

**31. **H. R. Gordon and A. Morel, *Remote Assessment of Ocean Color for Interpretation of
Satellite Visible Imagery, a Review; Lecture Notes on Coastal and Estuarine
Studies, Volume 4*. (Springer Verlag, New York,1983).

**32. **C. D. Mobley, *Light and Water: Radiative Transfer in Natural
Waters* (Academic Press, New York,1994).