Our retrieval method is based on a lookup table (LUT) approach [Christopher and Zhang, 2002; Wang et al., 2003]. In noncloudy conditions, the reflectance at the top of atmosphere (TOA) in the visible spectrum mainly is a function of Sun-satellite geometries (i.e., solar zenith angle θ0, viewing zenith angle θ, and relative azimuth angle ϕ), surface reflectance ρ0, AOT, and aerosol optical properties (AOP). Since the Sun-satellite geometry is known for each satellite pixel, and the reflectance of the ocean surface can be obtained through analysis of satellite data [Wang et al., 2003], the key aspect in satellite retrievals is to accurately model the aerosol optical properties that are primarily determined by the particle refractive index, size distribution and shape. There are at least 3 major unknowns in the retrieval algorithms, including the complex indices of refraction, size distribution and AOT. Traditional one-channel retrieval algorithms can only retrieve one parameter (e.g., AOT), whereas all the other parameters must be known priori [Mishchenko et al., 1999]. Multiple channels however can provide more information on aerosol spectral characteristics, and therefore can retrieve more aerosol parameters and improve the retrieval accuracy [Tanré et al., 1997; Higurashi and Nakajima, 2002, 1999; Mishchenko et al., 1999; Kaufman et al., 1997]. For example, in glint-free ocean scenes, the MODIS with its multi spectral capabilities simultaneously retrieves spectral AOT, effective radii and the ratio between different size modes (e.g., fine versus coarse) [Remer et al., 2002].
 The GMS5 has one visible channel and to retrieve AOT, aerosol optical properties must be properly characterized. Several studies have used a fixed (i.e., spatial-temporal independent) aerosol model [Wang et al., 2003; Zhang et al., 2001; Christopher and Zhang, 2002; Moulin et al., 1997; Ignatov et al., 1995] to infer the AOT information in an environment dominated by one aerosol type. However, due to the complexity of the aerosol in the east Asian region, it is neither sufficient nor reasonable to use a fixed aerosol model to calculate the aerosol optical properties. Therefore we have developed a dynamical aerosol model that can calculate the aerosol optical properties as a function of space and time. This model is implemented into the satellite retrieval algorithms to infer the AOT.
3.1. Dynamical Aerosol Model
 Remer and Kaufman  built a dynamic model for smoke aerosols in which aerosol properties (such as size distribution and phase function) varied as a function of aerosol optical thickness. However, this approach is not suitable for this study, since our goal is to retrieve the AOT from satellite observations. Therefore needed parameters from measurements other than GMS-5 must be determined and used to characterize the aerosol optical properties dynamically. Reid et al.  showed that the Ångström exponent α was well correlated with the aerosol size distributions, aerosol single scattering albedo and backscattering ratio, and proposed using α to estimate the variability in aerosol properties (α = −ln(τ1/τ2)/ln(λ1/λ2), where τ1 and τ2 are AOT at wavelength λ1 and λ2). The Ångström exponent α is also an important parameter in the two channel AHVRR retrieval algorithms [Mishchenko et al., 1999; Higurashi and Nakajima, 1999] and is closely related to relative importance of fine versus coarse aerosols in the aerosol size distributions [Tomasi et al., 1983]. Therefore this study employs Ångström exponent as an index to model the variability of aerosol optical properties.
 The aerosol size distribution used in this study is inferred from ground-based twin-scanning electrical mobility sizing (TSEMS) and optical particle counter (OPC) measurements at Gosan, Korea [Chun et al., 2001b; Brechtel and Buzorius, 2001]. Previous studies have demonstrated that the aerosol size distribution can be simulated by combining several lognormal size distributions [d'Almeida et al., 1991]. Figure 2 shows the measured daily mean as well as the 18-day mean volume size distribution at Gosan. Figure 2 also shows a bilognormal pattern in the measured size distribution. For a given day, particle volume distribution can be computed as:
where subscript n indicates the mode number; and rvn, σn and Cn are the volume median radius, standard deviation and the peak of nth mode, respectively. In this study, rvn value of 0.18 μm and 1.74 μm, σn of 2.16 and 1.78 were derived by fitting equation (1) to the 18-days mean size distribution (Figure 2) to represent the first and second mode, respectively.
Figure 2. Volume size distribution measured in Gosan, Korea, from 10 to 27 April 2001. Also shown is the simulated bilognormal size distribution.
Download figure to PowerPoint
 To use α as an index to model the variation of aerosol size distributions, a relationship between α and the size distribution must be established. Several studies have shown that α is closely related to the aerosol size distributions [e.g., Tomasi et al., 1983; Reid et al., 1999]. By assuming that aerosols have a Junge distribution dN/dln(r) = r−ν [Junge, 1955], we can show that α = ν − 2 [Liou, 2002]. However, the relationship between α and the size distribution may vary if aerosol size distribution does not follow a Junge size distribution [Tomasi et al., 1983]. Figure 2 shows that the measured aerosol size distribution has a distinct bilognormal pattern where the mode parameters (i.e., rvn, σn) have little day-to-day variations, but the mode peaks (C1 and C2) show day-to-day changes. It is therefore reasonable to use a peak ratio γ (defined as C2/C1 in equation (1)) to describe observed variations in aerosol size distribution. The mode peak ratio γ, which quantifies the relative abundance of two modes, has also been used for aerosol retrievals from satellites [e.g., Mishchenko et al., 1999; Higurashi and Nakajima,1999]. Higurashi and Nakajima  (hereinafter referred to as HN99) and Mishchenko et al.  used a bilognormal size distribution with rv1 of 0.17 μm and rv2 of 3.14 μm in their two-channel AVHRR global aerosol retrieval algorithms, and showed a polynomial relationship between γ and α. Both studies [Mishchenko et al., 1999; HN99] show that the relationship between γ and α can be established through Mie calculations if the parameters of two size mode (e.g., ri and σi) and refractive indices are known.
 The refractive index is highly variable depending on the chemical compositions of aerosols [d'Almeida et al., 1991]. Large differences in reported aerosol refractive indices also exist even for the same type of aerosol, for example dust [d'Almeida, 1987; Sokolik et al., 1993; Kaufman et al., 2001]. The aerosols in the study area are complex and are generally a mixture of several types of aerosols including sulfate, dust, sea salt and soot [Higurashi and Nakajima, 2002]. However the one-channel satellite retrieval algorithms characterize the column aerosol properties using a fixed effective refractive index [Rao et al., 1989; Ignatov et al., 1995; Wagener et al., 1997; Zhang et al., 2001; Christopher and Zhang, 2002; Wang et al., 2003]. (Effective refractive index does not refer to any specific aerosol type, but is suitable to quantify the composite radiative properties of all aerosols in an atmospheric column.) While the real part of the effective refractive index (i.e., 1.50–1.55) is consistent in the reported literature, the imaginary component of refractive index shows large variations. For example, the imaginary component of the refractive index at AVHRR ch1 wavelength (0.6 μm) varies from nonabsorbing [Rao et al., 1989; Wagener et al., 1997] to values considered absorbing 0.003–0.005i [Geogdzhayev et al., 2002; Mishchenko et al., 1999; HN99]. Recently a survey of aerosol properties from worldwide AERONET sites implied that the imaginary part of the refractive index of desert dust and oceanic aerosols range from 0.0015 to 0.0007 and the single scattering albedo varied from 0.95 to 0.98 at 0.67 μm [Dubovik et al., 2002]. Those single scattering values derived from Sun and sky measurements using a robust inversion technique [Dubovik et al., 2000; Dubovik and King, 2000], represent the effective aerosol properties in the atmospheric column and are suitable for satellite remote sensing retrievals algorithms. Dubovik et al.  found that dust aerosols have similar real part of refractive index when compared to values reported by Patterson et al.  for dust, while the imaginary part of refractive index was smaller, more consistent with the analysis of Kaufman et al . Although there is expected to be some amount of soot loading in our study area [Higurashi and Nakajima, 2002], the soot content and its effect on the column aerosol properties is still unknown. Therefore in this study, we use the real part of refractive index from Patterson et al.  but reduce the imaginary part refractive index by 70% to be consistent with AERONET retrievals [Dubovik et al., 2002]. We use the wavelength-dependent refractive index from Patterson et al.  in the radiative transfer model calculation to create the LUT (section 3.2). The refractive index used in this study at 0.67 μm is 1.53–0.002i.
 Using the above refractive index and the derived bilognormal distribution, we establish the relationship between α and γ through Mie calculations (Figure 3). Also shown in Figure 3 is the single scattering albedo (ω0) as a function of γ and the values used by [cf. HN99, Figure 3]. For the same γ, the ω0 values in the east Asian regions are larger than those in HN99 mainly because of the difference in the imaginary part of the refractive index (0.002i versus 0.005i). Using Figure 3, the Ångström exponent α can be used to derive γ, which can then be used together with derived size distribution mode parameters and refractive indices in Mie calculations to infer aerosol optical properties. In the HN99 global retrieval algorithms, the pair of (τ, α) is simultaneously retrieved from the two channel AVHRR algorithm by using a LUT in which TOA reflectance is a function of θ0, θ, ϕ, ρ0, τ, α. In this study, similar calculations are performed using a discrete ordinate radiative transfer model [Ricchiazzi et al., 1998] to create the LUT. However, since the GMS5 only has one visible channel, the (τ, α) pair cannot be retrieved simultaneously. Therefore, for each GMS5 pixel, the α values must be calculated from other sources. To achieve this, a successive correction method (SCM) is used to dynamically infer α from the ship, AERONET, and aircraft measurements (Appendix A). The SCM [Koch et al., 1983] is a relatively simple and widely used interpolation method that merges irregular point data from observation sites onto regular grids (Appendix A). In this study, we use this method to interpolate the Ångström exponent α inferred from the ground measurements (e.g., Sun photometers at different AERONET sites, Figure 1), ship measurement (e.g., Sun photometer on board NOAA R/V Ron Brown) and aircraft measurements (e.g., AATS6 on board C-130 and AATS14 on Twin Otter) into regularly spaced grids in the study area.
3.2. Retrieval Method
 The retrieval process has three major steps. Using the SCM technique, the first step is to create the daily spatial distribution of the Ångström exponent in the study region using ship, AERONET and aircraft measurements (Appendix A). The second step is to generate a background (clear sky) reflectance map and detect aerosols over the study area. Then the Ångström exponent is obtained for each aerosol pixel as identified by the GMS imager from step 1. The α value is then used to retrieve the AOT of each aerosol pixel from the previously computed LUT.
 This study uses the technique described in the work of Wang et al.  to derive the background ocean reflectance and to detect aerosol pixels. Using a minimum composite method, the spatial distribution of background or “clear sky” reflectance is obtained for each hourly GMS5 observation time [Wang et al., 2003; Zhang et al., 2001; Moulin et al., 1997]. Cloudy pixels are judged based on the IR temperature, spatial coherence (standard deviation) of the 3 × 3 pixel array in ch2 and ch3 images and the contrast of diurnal temperature (from infrared channels, ch2, and ch3) [Wang et al., 2003]. To further reduce cloud contamination, the spatial coherence of the 3 × 3 pixel array in ch1 images is used. Further details of this method can be found in the work of Wang et al . At this stage, for each aerosol pixel, the α value is available from the computed spatial distribution of Ångström exponent (Appendix A). The AOT is retrieved by finding the best match between the satellite reflectance and precalculated reflectance from the LUTs which is a function of θ0, θ, ϕ, ρ0, τ, α.