Simultaneous retrieval of aerosol optical thickness and chlorophyll concentration from multiwavelength measurement over East China Sea

Abstract

A flexible inversion algorithm is proposed for simultaneously retrieving aerosol optical thickness (AOT) and surface chlorophyll a (Chl) concentration from multiwavelength observation over the ocean. In this algorithm, forward radiation calculation is performed by an accurate coupled atmosphere‐ocean model with a comprehensive bio‐optical ocean module. Then, a full‐physical nonlinear optimization approximation approach is used to retrieve AOT and Chl. For AOT retrieval, a global three‐dimensional spectral radiation‐transport aerosol model is used as the a priori constraint to increase the retrieval accuracy of aerosol. To investigate the algorithm's availability, the retrieval experiment is conducted using simulated radiance data to demonstrate that the relative errors in simultaneously determining AOT and Chl can be mostly controlled to within 10% using multiwavelength and angle covering in and out of sunglint. Furthermore, the inversion results are assessed using the actual satellite observation data obtained from Cloud and Aerosol Imager (CAI)/Greenhouse gas Observation SATellite GOSAT and MODerate resolution Imaging Spectroradiometer (MODIS)/Aqua instruments through comparison to Aerosol Robotic Network (AERONET) aerosol and ocean color (OC) products over East China Sea. Both the retrieved AOT and Chl compare favorably to the reported AERONET values, particularly when using the CASE 2 ocean module in turbid water, even when the retrieval is performed in the presence of high aerosol loading and sunglint. Finally, the CAI and MODIS images are used to jointly retrieve the spatial distribution of AOT and Chl in comparison to the MODIS AOT and OC products.

1 Introduction

Aerosols are considered to exert considerable effects on global and regional climate change. They remain, however, one of the major uncertainty factors from the viewpoint of estimating and interpreting the Earth's energy budget and cloud owing to their high spatial and temporal variation and complex physical and chemical properties [Ramanathan et al., 2001; Li et al., 2011; Boucher et al., 2013]. It is also important to retrieve the optical properties of aerosols to improve the remote sensing of ocean color (OC), because the ocean body contributes only a small fraction of the signal at the top of atmosphere (TOA) where most information is obtained owing to backscattering of the molecular atmosphere and aerosols [Gordon and Morel, 1983]. Therefore, accurate aerosol estimation plays a significant role in determining ocean color, which is characterized by the optical properties of oceanic substances such as chlorophyll, inorganic suspended particles (suspended sediment), and colored dissolved organic matter (CDOM).

In the conventional ocean color remote sensing algorithms, the atmosphere and ocean systems are decoupled in two independent steps of atmospheric correction procedures to remove the influence of atmosphere and retrieval of the surface chlorophyll a (Chl) concentration [Gordon and Wang, 1994; Gordon, 1997; Fukushima et al., 1998; Antoine and Morel, 1999; Wang, 2010]. First, atmospheric correction is performed to estimate the influence of aerosols using sets of candidate aerosol modes, which are characterized by different optical properties and relative humidity values, to calculate radiance and compare it to the corresponding measurements in the red and near‐infrared (NIR) bands, where the ocean can be assumed black. Then, the best fitting aerosol modes are selected and extrapolated to shorter‐wavelength bands to calculate the water‐leaving radiance in those regions. Second, Chl is estimated from empirical or semiempirical relationships based on the statistical regression between remote sensing reflectance and oceanic substances [O'Reilly et al., 1998; Gordon et al., 1988; Carder et al., 1999; Lee, 2006; Hu et al., 2012].

There have been several other methods for improving the ocean color remote sensing, such as a linear combination index method by Frouin et al. [2006] to retrieve Chl based on the assumption that the measured reflectance combined in the visible and near‐infrared spectral bands can be related linearly to chlorophyll a concentration by a low‐order polynomial without explicit correction for aerosols. It has been proved that those approaches have good accuracy and availability for Chl retrieval and are employed in the operational processing of imagery from most ocean color sensors. Furthermore, under more challenging measurement conditions in the presence of nonnull reflectance in CASE 2 water or absorption aerosols, satellite‐derived normalized water‐leaving radiance is sometimes negative, and the assumption of zero water‐leaving radiance in NIR breaks down. Consequently, approaches and variants for improving classical atmospheric correction are conducted based on the assumption of spatial homogeneity of water reflectance or aerosols from nearby nonturbid areas or other practical methods [Hu et al., 2000; Ruddick et al., 2000; Pan and Mao, 2001; Wang, 2007; He et al., 2012; Mao et al., 2013].

To reduce the uncertainties discussed above, another method using the direct inversion algorithm, which is simultaneous retrieval of atmospheric and oceanic optical parameters, may be also a feasible way of complementing the prevailing schemes using the radiative transfer model to minimize the simulation output and measured reflectance [Chomko and Gordon, 1998; Doerffer and Fischer, 1994; Zhao and Nakajima, 1997; Stamnes et al., 2003; Xu et al., 2016] or Bayesian methodology to treat the likelihood of encountering specific values of marine reflectance as a probability distribution [Frouin and Pelletier, 2015]. These one‐step retrieval methods can be flexible to deal with the inversion problems both in CASE 1 and CASE 2 waters, or in the presence of absorption aerosol by defining the appropriate retrieved parameters and minimization algorithm to get convergence in an iterative manner. In this paper, an algorithm for simultaneous retrieval of AOT and Chl using multiwavelength measurement was developed. The forward radiation calculation was performed by a coupled atmosphere‐ocean radiative transfer model called Pstar [Ota et al., 2010] with an improved bio‐optical ocean module. The model uses the discrete ordinate and matrix operator method and was developed based on the Nakajima‐Tanaka scheme [Nakajima and Tanaka, 1983, 1986, 1988; Nakajima et al., 2000]. It has been proven to be highly accurate in simulating the radiation process both in the atmosphere and the ocean systems [Kokhanovsky et al., 2010; Shi et al., 2015]. The general structure of the model and the improved bio‐optical ocean module are introduced in section 2. Then, a nonlinear optimization estimation method [Rodgers, 2000] and design of the retrieval algorithm are shown in section 3. With these analyses, the accuracy of the proposed algorithm is investigated by numerical simulation; furthermore, the inversion results obtained from actual satellite data are compared to Aerosol Robotic Network (AERONET) AOT and OC products and the results of another retrieval algorithm that is used widely in satellite remote sensing.

2 Radiative Transfer Model in the Atmosphere‐Ocean System

2.1 Model Description

An accurate radiative transfer model is necessary for forward simulation of the radiation process in the atmosphere‐ocean system. Owing to the influence of the ocean surface, the source function in the radiative transfer equation (RTE) of the diffuse component of the Stokes vector is invoked by the direct solar beam reflected or transmitted by the sea surface, as shown below:

(1)

where τ is the optical depth measured from the top of the atmosphere (TOA), μ is the cosine of the viewing zenith angle measured from the negative z axis, and φ is the azimuth angle. The scalar ω denotes the single scattering albedo, defined as the ratio of the scattering coefficient to the extinction coefficient. P is the phase matrix and S is the source function for single scattering, which is different in the atmosphere and the ocean system [Ota et al., 2010]. In the atmosphere, it is defined as

(2)

and the ocean is assumed that no light is reflected from the ocean bottom, allowing S to be written as follows:

(3)

where F₀ is the solar flux vector, τ_a is the optical thickness of the entire atmosphere, and is the cosine of the refracted solar zenith angle in the ocean body. R_s and T_s denote the reflection and transmission matrixes of the ocean surface, which can be calculated by the Nakajima and Tanaka [1983] algorithm for rough ocean surfaces. The solution of the RTE is calculated based on the matrix and the adding methods in the inhomogeneous atmosphere and ocean layers [Nakajima and Tanaka, 1986; Ota et al., 2010] along with a highly accurate and efficient truncated method [Nakajima and Tanaka, 1988]. It has been proved that Pstar has a good accuracy in simulating the radiation process in both the atmosphere [Kokhanovsky et al., 2010] and the ocean body compared to the standard oceanic radiative transfer problems proposed by Mobley et al. [1993] [Shi et al., 2015].

2.2 Bio‐Optical Ocean Module

We introduce a detailed improved bio‐optical module in the model that Chl, suspended sediment, and CDOM are considered as depicted below.

The imaginary part of the refractive index of pure seawater was taken from the tabulated experimental data by Pope and Fry [1997] from 380 to 725 nm, in addition to empirical or experimental data for other wavelengths [Hale and Querry, 1973; Smith and Baker, 1981; Kou et al., 1993]. The absorption coefficient was calculated using a_w = 4πn_i(λ)/λ, where λ is the wavelength and n_i(λ) is the imaginary part of the refractive index. The real part of seawater refractive index was taken from Hale and Querry [1973] for longer wavelengths and calculated using Huibers [1997] formulation for 300–800 nm range, which was derived from Quan and Fry's [1995] expression:

(4)

where n₀ = 1.31405, n₁ = 1.779e‐4, n₂ = −1.05e‐6, n₃ = 1.6e‐8, n₄ = −2.02e‐6, n₅ = 15.868, n₆ = 0.01155, n₇ = −0.00423, n₈ = −4382, and n₉ = 1.1455e6. T is the temperature in °C, S is the salinity in PSUs (practical salinity units), and n_a is the refractive index of air which was calculated according to Zhang and Hu [2009]. It is studied that large changes in temperature and salinity impact seawater absorption and scattering coefficient at specific wavelengths [Hollstein and Fischer, 2012]. Hence, the influence of temperature and salinity was both considered based on the research of Zhang et al. [2009] and Röttgers et al. [2010] as follows:

(5)

(6)

where a_w and b_w are the absorption and scattering coefficients, which are related to wavelength, temperature, and salinity; ψ_T(λ) and ψ_S(λ) are the temperature and salinity correction coefficients for water absorption; β_w(90^∘, T, S, λ) is the volume scattering function at 90° as a function of wavelength, temperature, and salinity; δ_w is the depolarization factor, and it is defined as 0.039 by default [Zhang et al., 2009]. Other elements of the seawater phase matrix were corrected based on the ratio of the Rayleigh scattering matrix to the first parameter.

The living phytoplankton strongly absorbs visible light, and it is the principal component in CASE 1 waters. The absorption coefficient of phytoplankton was computed using Bricaud et al.'s [1995] empirical formula:

(7)

Here A(λ) and B(λ) are positive principal wavelength‐dependent parameters. We used the compiled values by Mobley [2014], where A(λ) and B(λ) of Bricaud et al. [1995] in a spectral range between 400 and 700 nm were integrated with the data from Vasilkov et al. [2005] and Morrison and Nelson [2004] to cover a wider wavelength range.

The scattering coefficient of phytoplankton b_ph(λ) and the backscattering fraction B_bp were calculated according to Huot et al.'s [2008] method, which is based on Morel and Maritorena [2001]:

(8)

The size distribution of ocean particles is often found to follow the Junge distribution [Stramski and Kiefer, 1991]. The wavelength dependence of the particle phase function is neglected, which is similar to the research of Fischer and Grassl [1984]. In this study, the Fournier‐Forland (FF) phase function [Fournier and Forand, 1994; Fournier and Jonasz, 1999] was used to represent the scattering angular distribution of phytoplankton. FF is an analytical approximation form of the phase function of an ensemble of particles that have a hyperbolic size distribution based on the exact Mie theory calculation, and it has been proved to be more realistic than other analytical model developed thus far [Mobley et al., 2002]. FF is written as follows:

(9)

where

(10)

Here n is the real part of the particles' refractive index, μ is the slope parameter of the Junge distribution, Θ is the scattering angle, and δ₁₈₀ is the value of δ at 180°. Based on the FF function, the backscattering fraction can be determined analytically as well [Mobley et al., 2002]:

(11)

Furthermore, Mobley et al. [2002] proposed a simple linear relationship between μ and n to reproduce the FF fit to the observation from Petzold [1972] data, as expressed below:

(12)

Since the value of μ varied in the majority of cases between 3.0 and 5.0 [Jonasz, 1983; Chowdhary et al., 2006], a look‐up table was generated to relate B_{bp − FF}, μ and n using equations 10-12. Based on the best fits of B_bp and B_{bp − FF} for a given Chl, the values of μ and n can be determined to calculate the FF function using equation 9, which is similar to the equations of Zhai et al. [2010] and Xu et al. [2016].

The yellow substance absorption coefficient a_y(λ) related to the chlorophyll concentration was calculated using Morel's [2009] formulation:

(13)

where λ₀ is set to 400 nm and a_y(400, [Chl]) = 0.065[Chl]^0.63; the scattering coefficient was set to zero because the yellow substance has only the absorption property. Based on the discussed above, the bio‐optical oceanic module for CASE 1 water that all inherent optical properties are determined by the value of Chl was developed.

In order to extend the application in practice, we tried a more complex bio‐optical module for the simultaneous retrieval of AOT and Chl in CASE 2 water. The influence of sediment and independent CDOM is considered as well, of which the optical properties were calculated using the empirical formulas of Tassan [1994] and Bricaud et al. [1998]:

(14)

(15)

where b_sed(λ) is the scattering coefficient of sediment (m⁻¹), S is the suspended sediment concentration (g m⁻³), b_s(550) = 1 m² g^− 1, and n_s = − 1. The absorption coefficient of sediment is assumed to zero, backscattering ratio is defined as 0.015, and a_y(λ) is the absorption coefficient of CDOM (m⁻¹).

The TOA radiance is affected by the reflection from the ocean as Figure 1 suggested that the reflectance in 0.44 µm at TOA between different oceanic depth values for the same Chl or CDOM could have obviously difference in low concentration conditions. In the current algorithm, we assume the ocean is deep enough to neglect reflection from the ocean bottom. Based on this assumption, we set the ocean depth at the euphotic depth Z_e, which is defined as the depth at which the downwelling irradiance is reduced to 1% of its value at the surface based on the formulations of Morel and Berthon [1989] from Lee et al. [2007], as follows:

(16)

Figure 1

Reflectance at TOA of 0.44 µm and 0.55 µm at different substance concentrations and ocean depths. The AOT of 0.2 at 550 nm and no ocean bottom reflection are assumed.

The depth of surface layer, which is defined as having a thickness corresponding to one attenuation length, is defined as Z_pd ≈ Z_e/4.6 [Morel and Berthon, 1989] which describes the surface oceanic depth where surface chlorophyll concentration exists.

The total chlorophyll concentration was calculated as follows [Morel and Berthon, 1989]:

(17)

The vertical distribution of chlorophyll was assumed to be Gaussian, where each layer of chlorophyll concentration [Chl]_layer(z) was defined as follows:

(18)

where C_b, C_max, ξ, ξ_max, and Δξ were calculated according to Morel and Berthon's [1989] method.

3 Optimization Method for Ocean Color Retrieval

Many optimal estimation methods have been proposed for the nonlinear retrieval problem. In general, the measurement vector y, such as observed radiance or reflectance in each wavelength, is expressed by a forward radiative transfer model F(x, b) with error ε [Rodgers, 2000]:

(19)

where F(x, b) describes the complete physics of the measurement, x is the state vector to be retrieved, b comprises the quantities that influence the measurement but are not to be retrieved, and ε is the error vector consisting of measurement error and forward model error. The probability density function (pdf) of the state vector given the measurement can be determined from Bayesian theory:

(20)

where all pdfs are assumed to be Gaussian distributions. The posteriori pdf is expressed as the following cost function in the logarithmic scale:

(21)

Here x_a denotes the a priori values of the state vector, S_ε is the measurement error covariance matrix, and S_a is the variance‐covariance matrix defined by a priori state values. In the maximum a posteriori (MAP) approach, x is selected to find the zero of the gradient of J(x) so that the posteriori pdf becomes maximum under the condition that a priori information is already given [Rodgers, 2000]. Then, the Newton iteration method optimized by the Levenberg‐Marquardt algorithm [Levenberg, 1944; Marquardt, 1963] is used as follows:

(22)

(23)

where subscript i denotes the index of the ith iteration, K is the Jacobian matrix or weighting function, which is derived by the forward model to the state vector as K = ∂F(x, b)/∂x, γ is a nonnegative parameter chosen in each iteration to minimize the cost function using the method of Press et al. [2007], and A is the averaging kernel matrix that represents the sensitivity of the retrieval to the true state.

Aerosol determination with the remote sensing algorithm depends primarily on the aerosol model used, which should be interpreted as representatively as possible [Harmel and Chami, 2011]. In the current algorithm, a scattering multicomponent approach was adopted in the aerosol model. In this method, instead of using the average refractive index of a single component to represent a mixture of different types of aerosols, the multicomponent approach, in which the refractive index of each type of aerosol particle is calculated considering hygroscopic growth [Shettle and Fenn, 1979], was used. This approach has been proved to be more realistic in the atmospheric correction of ocean color imagery [Yan et al., 2002]. It is assumed that the external mixture aerosol model consists of fine particles, sea salt particles, and dust/yellow sand particles defined as nonspherical aerosol particles; an internal mixture of urban aerosol and soot exists in the fine particles. In this study, the phase matrix of nonspherical aerosols was calculated using Dubovik et al.'s [2002b] method. For each particle mode, the size distribution is assumed to follow lognormal, which is reported to be a good approximation of observations [Nakajima and Higurashi, 1997; Higurashi et al., 2000; Dubovik and King, 2000; Dubovik et al., 2002a] as follows:

(24)

where V is the aerosol volume density, r_m is the mode radius, and ln s is the standard deviation of ln r. It was studied that the retrieval uncertainties depend much more strongly on AOT than on the other optical properties of aerosols based on the analysis of information content over ocean [Knobelspiesse et al., 2012]. Therefore, the size distribution parameters of each particle were fixed and defined in the current algorithm, as summarized in Table 1.

Table 1. Aerosol Size Distribution Parameters in This Study

Type	rm(µm)	s
Fine	0.12	1.68
Sea salt	2.2	2.01
Dust/Yellow sand	4.0	3.0

In addition to the parameters describing the optical properties of aerosols, the coupled ocean parameters need to be determined. The reflectance and transmission functions of ocean surface were calculated using the method of Nakajima and Tanaka [1983] for the rough ocean surface case, and these functions were driven by wind speed similar to the Cox and Munk's ocean model [Cox and Munk, 1954] but without offset. For CASE 1 water, in which phytoplankton is the dominant substance influencing the ocean's optical property, the state vector x consists of six parameters: AOT of fine particles, AOT of sea salt, AOT of dust/yellow sand, volume soot fraction in fine particles, wind speed, and Chl. For CASE 2 water, x contains eight parameters after the addition of sediment and CDOM. The soot fraction in fine particles is defined as the retrieval parameter owing to its high absorption effects. To perform better retrieval, a global three‐dimensional spectral radiation‐transport aerosol model called SPRINTARS [Takemura et al., 2000] was used to generate the a priori values of AOT for each particle because SPRINTARS has been proven to accurately predict aerosol radiative forcing and climate effects [Takemura et al., 2002; Myhre et al., 2013]. Other supporting information such as atmospheric humidity profile, ocean temperature, and pressure were taken from National Centers for Environmental Prediction (NCEP) reanalysis data. The atmospheric and oceanic were divided into six and four layers, respectively. Clear‐sky pixels were selected using Ishida and Nakajima's [2009] cloud detection algorithm. The general flowchart of the retrieval algorithm is shown in Figure 2.

Figure 2

Flow chart of retrieval algorithm.

4 Validation and Discussion

4.1 Simulation Results

To investigate the accuracy and stability of the proposed algorithm, a numerical simulation was performed before using actual satellite data. In the current algorithm, only radiance information is used. For all retrievals, synthetic measurements (input values) of radiance for a given atmospheric state and random state vector values were generated at six wavelengths (380, 490, 550, 670, 765, and 865 nm) and three observation angles (0°, sunglint and backscattering) using the radiative transfer model, the total number of simulated data is 200. It is noted that the sunglint observation was used because it generates robust information in the retrieval of wind speed and is beneficial from the viewpoint of determining absorption aerosols as a bright background [Kaufman et al., 2002]. Then, simulated retrievals (output values) were performed based on the optimal estimation theory described in section 3. To represent a more realistic circumstance, we added a random error to the radiance measurement, which is a common strategy to test the influence of measurement error on the performance of remote sensing algorithms, i.e., a relative measurement uncertainty of ±2% to the radiance. Figure 3 shows the results of the numerical simulation conducted to simultaneously retrieve AOTs and chlorophyll concentration. It is demonstrated that the relative errors in jointly determining the AOTs of fine and sea salt particles were both less than approximately 10%. In this regard, it is found that AOT of sea salt particles with lager geometry properties are more accurately estimated compared with that of fine aerosols. It is important to note that coarse particles including sea salt and dust aerosol could be distinguished approximately, especially for the determination of nonabsorption sea salt particles, although under low dust loading condition, the amount of dust particles is slightly overestimated compared to its “true value.” We guess this successful separation is attributed to using 380 nm at which dust particle absorption is significant. The volume soot fraction in fine particles was difficult to retrieve owing to its feeble sensitivity to the measurement (not shown). In the ocean, the relative error in the inversion of wind velocity can be estimated with high accuracy for observations made in sunglint regions (not shown), which is similar to the research of Harmel and Chami [2012] using polarized information. The chlorophyll concentration could generally be retrieved with the relative error controlled to 10%, even though slight overestimation and underestimation occur under very low and high Chl, respectively. Statistics of the total degree of freedom of signals (trace of averaging kernel matrix) were mostly over 5.0, which are comparable to the number of state parameters and revealed that the retrieval was stable and not seriously affected by measurement noise. It is noted that the Chl can be retrieved using multi‐information observation even contaminated by the sunglint signal, which may ascribe to the better estimation of aerosol and wind speed to generate more accurate forward simulation at TOA using radiative transfer model than the conventional linearly equation added from various distinct physical contributions in the sunglint region. However, the retrieval simulation using single observation angle reveals the accuracies of AOTs of sea salt, dust, and Chl are decreased in sunglint conditions compared with those using out‐of‐sunglint observation.

Figure 3

Simulation of simultaneous retrieval of AOTs for fine, sea salt, and dust particles at 550 nm, and chlorophyll a concentration with input values. The blue lines denote the values 10% smaller or larger than the input values.

4.2 Retrieval Results Using Actual Satellite Data

4.2.1 Using GOSAT/CAI Radiance Data Over East China Sea

Following the simulation retrieval experiment performed using the radiance data generated by the model discussed above, actual satellite observation data were used to further assess the algorithm's availability. To this end, observations obtained using the Greenhouse gas Observation SATellite (GOSAT) launched in January 2009 were analyzed, and GOSAT is designed to measure carbon dioxide loading using the Thermal And Near infrared Sensor for carbon Observation Fourier Transform Spectrometer [Yokota et al., 2004]; in addition, the satellite carries the Cloud and Aerosol Imager (TANSO‐CAI) with four bands (380, 674, 870, and 1600 nm) for cloud screening and aerosol detection. Radiometric correction was conducted as prescribed in Shiomi et al. [2010]. In this study, we collected GOSAT/CAI level 1 data and AERONET level 2.0 aerosol and ocean color products from Gageocho and Ieodo stations between 2012 and 2014 [Holben et al., 1998; Zibordi et al., 2009] over East China Sea. The two aforementioned sites are located at 33.942°N, 124.593°E and 32.123°N, 125.182°E, respectively. It is noted that only the data that satisfied the following criteria were used for the comparison: (1) more than 20 satellite pixels should be analyzable successfully in a 5 × 5 pixel window around each site; (2) AERONET aerosol and ocean color products should be observed on the same day; and (3) the time difference between AERONET aerosol observation and satellite overpass should be less than 30 min.

Figures 4a and 4b show the comparison of satellite simultaneous retrieved AOT at 550 nm and Chl using CASE 1 module from CAI with the AERONET observation at Gageocho site. Here AOT refers to the total aerosol optical thickness, which is the sum of the optical thickness of fine, sea salt, and dust particles. It is demonstrated that the AOT is underestimated in several cases compared with the values reported by AERONET. For chlorophyll concentration retrieval, the result is generally comparable to values from AERONET, although there is one UV wavelength of 380 nm used to Chl retrieval. However, there are several overestimation occurred in a few cases during the retrieval. Another comparison was performed at the Ieodo site, as shown in Figures 4c and 4d. The AOT retrieval results are better than those obtained at the Gageocho site, and the retrieved AOT mostly compares favorably to AERONET reported values, with the exception of two obvious overestimation cases on 3 December 2013 and 29 January 2014, as shown by green and white circles, respectively, of which the optical thickness of dust particles is overestimated extraordinarily. The correlation coefficient is greater than 0.9; root‐mean‐square error is 0.0864. As long as chlorophyll concentration, the values from AERONET are constant and show less temporal variation compared with those obtained from the Gageocho site in spite of the two sites being close, the retrieval results are generally overestimated compared with AERONET, partly due to the underestimation of CDOM, which significantly affects light absorption at 380 nm. Therefore, a correction for CDOM using CASE 2 ocean module is performed for the simultaneous retrieval of aerosol and Chl. It is demonstrated that the accuracies of retrieved Chl are mostly improved compared with those obtained using CASE 1 water module shown in Figure 5d. However, there are not significant variations for the retrieval of AOT as suggested in Figures 5a and 5c, which may be caused by uncertainties estimation of the volume soot fraction in fine particles.

Figure 4

Comparison of satellite simultaneously retrieved AOT and Chl (mg m⁻³) from CAI with AERONET observations at the Gageocho ((a) 550 nm AOT; (b): Chl) and Ieodo (c) 550 nm AOT; (d) Chl) sites using CASE 1 water module. The simultaneous retrieved results of AOT and Chl are shown in same color.

Figure 5

Similar to Figure 4 but for retrieval using CASE 2 ocean module.

4.2.2 Application to CAI Imagery

The retrieval experiment was then conducted using CAI level 1b images obtained around the East China Sea on 13 March 2012. Spatial distribution with low resolution of the retrieved total AOT and Chl are shown in Figure 6. The acquired image was selected because it shows relatively obvious variability in both aerosol and chlorophyll concentration throughout the entire scene. For AOT retrieval, high AOT values (0.8~) was observed near the eastern coastal area of China, and relative low AOT loading (~0.1) was located at the central of Yellow Sea. In regard to Chl retrieval, high Chl values (6mg m⁻³~) can be seen in the party ocean area of Bohai and west coastal of Korea, while there are low chlorophyll concentration in the middle of Yellow Sea and south ocean area of South Korea, the general spatial distribution of AOT and Chl is similar to those generated from MODerate resolution Imaging Spectroradiometer (MODIS)/Aqua products (not shown).

Figure 6

Spatial distribution of (a) simultaneous retrieved total aerosol optical thickness and (b) chlorophyll concentration using CAI imagery on 13 March 2012 over East China Sea.

4.2.3 Using MODIS Aqua Data

To further investigate the availability of the proposed algorithm, MODIS Level 1B‐calibrated reflectance data with a greater number of spectral bands of 8 wavelengths (412, 442, 487, 554, 670, 746, 867, and 1620 nm) were used to jointly retrieve AOT and Chl. Data exclusion criteria were similar to those in section 4.2.1. The retrieval experiment was conducted using CASE 1 water module first. It is demonstrated that the AOT values retrieved from the Gageocho and Ieodo stations are both overestimated by comparing the results with AERONET observation (Figures 7a and 7c), even when using the data of a greater number of spectral bands compared with GOSAT/CAI. This overestimation may be attributed to the fact that the sediment in the ocean during the retrieval process was ignored because reflectance in the 0.55, 0.66, and 0.86 µm channels was affected by the presence of sediment [Li et al., 2003] since the East China Sea was always assumed as typical CASE 2 water [Matsushita et al., 2012], and the spectral reflectance does not fit very well with a power law formula. The root‐mean‐square errors in AOT at the Gaogecho and the Ieodo sites are 0.0925 and 0.105, respectively. Chlorophyll concentration retrieval was more accurate at the Gaogecho site than that at the Ieodo site, as shown in Figures 7b and 7d, respectively, of which the root‐mean‐square errors are 1.08 and 1.51, respectively. Moreover, AOT and Chl can be retrieved even when the satellite observation covers a sunglint region where the reflected Sun angle lies between 0° and 36° [Ackerman et al., 1998], shown as triangle, in Figure 7. A comparison between Figures 7b and 7d shows that the chlorophyll concentrations retrieved at the Ieodo site (Figure 7d) are generally overestimated, which are similar to the results obtained using GOSAT/CAI satellite data, as in Figure 4d, which might also be caused by CDOM underestimation in the retrieval. Notably, owing to different satellite overpass times and cloud conditions between the two sets of satellite data, despite being of the same color, Figures 4 and 7 do not show the same results.

Figure 7

As in Figure 4 but for retrieval using MODIS Aqua data. The triangle means the observation is covered in sunglint region.

Due to the general overestimation of AOT and chlorophyll concentration, a correction for sediment and CDOM in the retrieval process is conducted. Figure 8 shows the retrieval results obtained using the CASE 2 ocean module and eight‐channel MODIS reflectance data, which are the same as those above. It is demonstrated that the retrieval results of both AOT and Chl are improved significantly compared to those obtained using the CASE 1 water module, as in Figure 7. At the Gegeocho station, the root‐mean‐square error decreased to 0.0307 and 0.839 for retrieval of AOT and Chl. A comparison of Figures 7a and 8a shows that the AOT accuracy is improved significantly after considering the contamination of sediment. Moreover, the retrieved results of Chl are more consistent to the AERONET reported values, even though the observation is covered in sunglint region as denoted by the triangle (Figure 8b). At the Ieodo study site, the root‐mean‐square error of the retrieved AOT and Chl decreased to 0.0509 and 0.447. Especially, the retrieval results of chlorophyll concentration compared more favorably to AERONET reported values, as shown in Figure 8d. Furthermore, in the case of high aerosol loading, denoted by the blue circle, the AERONET site reported a high 550 nm AOT of 0.612 and Chl of 2.396 at 05:23 UTC and 04:26 UTC on 29 January 2014, respectively, and the retrieval for the MODIS/Aqua data acquired at 05:00 UTC gave accurate values of 0.613 and 2.148 for AOT and Chl, respectively. In order to make a more sufficient estimation of algorithm accuracy, more matchups available data from USC_SEAPRISM site are used. It is demonstrated that the retrieved AOT and Chl are also well compared with AERONET reported values as shown in Figure 9, with the correlation coefficient up to 0.9212 and 0.8681, respectively.

Figure 8

Similar to Figure 7 but for retrieval using CASE 2 ocean module.

Figure 9

Similar to Figure 8 but for the retrieval in USC_SEAPRISM site.

4.2.4 Application to MODIS Aqua Imagery

The retrieval algorithm was applied to selected MODIS level 1b images obtained around the East China Sea on 18 October 2011 with low resolution interpolated from the original 1 km resolution image. Spatial distributions of the simultaneous retrieval of total AOT, AOT of fine particles, and Chl obtained from the observation of eight wavelengths are shown in Figures 10b, 10d, and 10f. The MODIS Level 3 gridded AOT and Standard Mapped Image (SMI) Chl products with the same resolution obtained on the same day, shown in Figures 10a, 10c, and 10e, are used for comparison. For AOT retrieval, high aerosol loading (0.6~) was observed near the eastern coastal area of China, and relatively low AOT values (~0.1) are retrieved in the middle sea of China and Korea, and the western coastal area of Korea, as shown in Figure 10b. In general, the retrieval results are mostly similar to those obtained using MODIS products with the correlation coefficient up to 0.9438, as shown in Figure 10a, while the retrieved AOT values are slightly lower than those obtained using the MODIS aerosol algorithm on the southwest coast of Japan. This difference is caused by the lower contribution of fine aerosols to the total optical thickness, as accounted for in our algorithm, than those values achieved from MODIS AOT products, as shown in Figures 10c and 10d. A comparison of the AOT of fine particles shows that the retrieval results of our algorithm are generally lower than those of MODIS products which have been reported an overestimation of fine mode fraction compared with AERONET results [Kleidman et al., 2005], and this difference might be explained by the different estimations of sediment or different size parameters used in the two algorithms. Figure 10f demonstrates the simultaneously retrieved results of chlorophyll concentration using our algorithm. High Chl values (6mg m⁻³~) can be seen near Korean Bay, Bohai Bay, and the west coast of South Korea, whereas relatively lower values (~0.5 mg m⁻³) were retrieved in the Southern Ocean area of South Korea and southwestern ocean area of Japan. The general patterns of simultaneously retrieved chlorophyll concentration are also similar to those generated from MODIS OC products, as shown in Figure 10e. However, the Chl values are significantly lower than those retrieved by the MODIS OC algorithm near the eastern coastal area of South Korea, where our results are more similar to those of MERIS SMI products (not shown). Another obvious discrepancy is located at the center of the Yellow Sea, where the simultaneously retrieved Chl values are slightly larger than those obtained using MODIS OC products. This may be caused by the different estimations of absorptive aerosol or CDOM between our method and the MODIS algorithm. Statistic of histograms of the frequency values of AOT and Chl from the two types of algorithm are shown in Figure 11, in general, there are not significant differences of retrieved total AOT between those two algorithms; however, the retrieved AOT of fine particles is smaller than those from MODIS products. The Chl values in this study are overestimated in relatively low Chl conditions and underestimated in high Chl conditions compared with those retrieved from MODIS algorithm.

Figure 10

Level 3 satellite images acquired on 18 October 2011 over East China Sea. (a) MODIS total AOT at 550 nm, (c) MODIS fine AOT at 550 nm, and (e) MODIS chlorophyll concentration in mg m⁻³. Simultaneously retrieved results using the algorithm developed in this paper. (b) Total aerosol optical thickness at 550 nm, (d) Fine aerosol optical thickness at 550 nm, and (f) chlorophyll concentration.

Figure 11

Statistic of histograms of the frequency values of (a) total AOT, (b) fine AOT, and (c) Chl from the two types of algorithm.

5 Conclusions

In this paper, a flexible full‐physical retrieval algorithm was developed to simultaneously determine aerosol optical thickness and chlorophyll concentration based on multiwavelength measurements. It is different from the conventional ocean color algorithms in its use of a two‐step method and decoupling of the atmosphere‐ocean system. The current algorithm uses an accurate radiative transfer model to simulate the forward radiation process in the atmosphere‐ocean system based on an updated bio‐optical module, influence of ocean temperature and salinity, new chlorophyll concentration‐related data set ranging from 0.3 to 1 µm, and a more comprehensive optical model are considered in the bio‐optical ocean module. Moreover, a global aerosol transport‐radiation model was used as the a priori constraint to increase the aerosol retrieval accuracy. To investigate the algorithm robustness to noise, a simulation experiment with 2% random error was first performed. The relative errors in determining AOT and Chl simultaneously are less than approximately 10%. Moreover, case studies using actual satellite data from GOSAT/CAI and MODIS Aqua demonstrate that the simultaneously retrieved AOT and Chl compare favorably to the results of AERONET AOT/OC products, even in sunglint region and the high aerosol loading cases. Finally, the CAI and MODIS images were used to simultaneously retrieve the spatial distribution of AOT and Chl; the results shows good consistence of retrieved total AOT between those two algorithms in general; however, some differences are demonstrated that the retrieved AOT of fine particles are smaller than those from MODIS products. The Chl values in this study are overestimated in relatively low Chl condition and underestimated in high Chl condition compared with those retrieved from MODIS algorithm.

The advantage of this algorithm lies in its feasibility to deal with CASE 1 and CASE 2 water and also the situation in the presence of absorption aerosol. However, a main drawback of the current algorithm is the slow convergence in some cases that constrain its application in large‐scale satellite data. Moreover, several cases contaminated by sunglint could be retrieved in this study, which may ascribe to the better estimation of fine aerosol and wind speed to generate accurate forward simulation at TOA using radiative transfer model, while it might introduce obvious error in the presence of high sea salt loading and more complex ocean circumstance. In future, validation and improvement of the algorithm performance using additional in situ data and the multipixel method are required. Moreover, development of an acceleration algorithm using look‐up table [Higurashi and Nakajima, 1999] or networks [Takenaka et al., 2011] method is also a part of our future work.

Acknowledgments

This work was supported by funds from MOEJandJAXA/GOSATandGOSAT2, JST/CREST/EMS/TEEDDA, JAXA/EarthCAREandGCOM‐C, MEXT/Kakenhi/Innovative Areas 2409, MOEJ/ERTDF/S‐12, Key Laboratory of Meteorological Disaster of Ministry of Education, and Nanjing University of Information Science and Technology (KLME1509). The authors express their sincere thanks to the relevant PIs for establishing and maintaining the AERONET AOT/OC sites used in this investigation, GOSAT, MODIS, and NCEP science teams. GOSAT/TANSO‐CAI data are provided by JAXA/EORC. MODIS/Aqua data are provided from NASA website at https://ladsweb.nascom.nasa.gov/data/search.html. NCEP reanalysis data are achieved at http://www.esrl.noaa.gov/psd/data/gridded/. The authors are grateful to Yoshifumi Ota, Hiroshi Murakami, and Hideaki Takenaka for useful discussions.