From equation (1), we can see that each term on the right hand side is a function of AOD. Quality of the surface reflectance estimates is strongly driven by the knowledge of AOD. Thus, the accurate retrieval of aerosols is an important step. When the reflectance of surface is low, the scattering effect dominates over dark surfaces, while this effect is mixed with absorption effect over brighter surfaces. Therefore, the aerosol scattering effect is stronger for low surface reflectance, and could be used in an inversion procedure to assess the AOD. Kaufman and Sendra (1997) suggested using dense dark vegetation (DDV), usually green forests, as the dark targets to derive the AOD. In order to estimate AOD over land, the sur-face reflectance of these dark pixels must be estimated to an uncertainty of 0.005-0.01. This uncertainty translates to an error in optical depth of 0.05-0.1 (Kaufman, 1993). Therefore we need a method to assess the surface reflectance through an aerosol layer of unknown op- tical depth. There are three methods often used to estimate the reflectance of the land surface. Kaufman and Sendra (1997b) used a vegetation index measured at TOA to detect the dense dark vegetation and then gave the reflectance. Since the vegetation index is affected significantly by the aerosol (Higurashi et al., 1999), the threshold value that defines dense vegetation varies from image by image depending on the aerosol concentration. To solve this problem, Holben et al., (2001) and Kaufman et al. (1988), used the 3.75 µm channel to de- tect dense vegetation. For longer wavelength, the reflectance is slightly affected by aerosol (Kaufman et al., 1988), except dust (Berk et al., 1989). Moreover, the 2.1 µm channel is presented on Global Imager GLI, EOS-MODIS and Landsat TM. The advantage of the 2.1 µm channel over the 3.75 µm channel is that it is not affected by emitted radiation. The 2.1 µm channel is almost transparent to most aerosol types (except dust) and therefore can be used to detect dark surface targets. Correlation between the surface reflection in the blue (0.49 µm), red (0.66 µm), and 2.1 µm has been established (Kaufman, 1988). Results from a variety of surface covers show that the surface reflectance at 0.49 µm ( ) and 0.66 µm can be predicted from that at 2.1 µm In the case of less than 0.1, one can obtain and It is a challenge to retrieve AOD from HJ-1A/B CCD data due to lack of spectral band around 2.1 µm. Although aerosol loadings can be obtained over some regions from vegetation indices, such as NDVI, the correction accuracy is still greatly affected by aerosol and the determination of DDV regions with great subjectivity (Kaufman et al., 1997b). The method proposed by Sun et al. (2011) is used to retrieve aerosol loadings, which uses MODIS surface reflectance output (MOD09) as the priori datasets to support HJ-1A/B CCD aerosol retrieval. The spatial resolution of HJ-1A/B CCD was transferred to 500 m and the correlation function of HJ-1A/B CCD and the MOIDS apparent reflectance were built based on their spectral response function and the spectrum of typical land types. MOD09 is used to provide the percentage of dense vegetation distributions, and then AOD is retrieved from equation (1) based on the surface reflectance relationship between red and blue bands over DDV regions developed by Kaufman (1988).

The surface reflectance can be estimated by solving the radiative transfer (RT) equation. To save time, here, we estimate surface reflectance with look-up table (LUT). This LUT was built by Second Simulation of the Satellite Signal in the Solar Spectrum (6S). 6S TR code is used for simulating the satellite signal between 0.25 μm and 4.0 μm (Vermote, 2006). Absorption and scattering of atmospheric gases and multi-scattering of aerosols are considered in calculating atmospheric transmittance. Users can measure surface data to define their own atmospheric model or use six different atmospheric models defined in 6S. Standard aerosol models, including continental model, maritime model, urban model, the Shettle model for background desert aerosol, biomass burning, and a stratospheric model are defined in this code. Users can also use their own aerosol model.

A LUT for surface reflectance can be produced through calculating the relationships among surface reflectance and main atmospheric parameters L₀, S and transmittance under different AOD based on the settings of relevant parameters (Vermote, 2006). The following parameters were considered: (1) Observation conditions including solar zenith, solar azimuth, sensor zenith and sensor azimuth angles. Six sensor zenith angles from 0° to 35° with a step of 6°; 18 solar zenith angles from 0° to 85° with a step of 5°; 19 relative azimuth angles from 0° to 180° with a step of 10°. (2) Atmospheric model, northern mid-latitude summer and mid-latitude winter atmospheric models were selected in this LUT constructing. (3) Aerosol model, four aerosol models (the continental, oceanic, dust and biomass burning aerosol models) with AOD at 550 nm in 10 grades: including 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.8, 1.0, 1.5 and 2.0 are used to describe the aerosol conditions.

Six images in three different areas representing typical areas of China are selected in this study: Nanjing in Jiangsu province with high vegetation coverage and serious air pollution; Yuzhong in Gansu province with low vegetation coverage and much air pollution; Taiwan with dense vegetation coverage and relatively clean air. The radiometric calibration and the geometric correction of the images have been performed. In order to illustrate the aerosol optical depth more clearly, we cut out the part images from the six HJ-1A/1B CCD images with the administrative boundary line of Nanjing, Jiangsu province, Yuzhong, Gansu province and Taiwan. These images are very clear except for a small patch of cloud in Nanjing City and some parts of Taiwan. In the false color composite image, the red pixels are covered by vegetation. We can see that most parts of Nanjing and Taiwan are covered by dense vegetation, while only the southern part of Yuzhong is covered by vegetation.

With the DDV data obtained from MOD09 data, AOD is retrieved from the six HJ-1 CCD images. Figure 1 is the results and we can see that the aerosol distributions in Nanjing and Taiwan are denser than that in Yuzhong of Gansu. This can be explained by the fact that ar-eas covered by dense vegetation in Nanjing and Taiwan are bigger than the area in Yuzhong. Some ground-based in-situ measurements of AOD acquired from AERONET (AErosol RObotic NETwork) is used to validate the estimated AOD results. The AERONET is a eries of ground-based remote sensing aerosol networks established by NASA and PHOTONS. The network provides a long-term, continuous and readily accessible public database of aerosol optical, microphysical and radiative properties for aerosol research and characterize-tion, as well as validation of satellite retrievals. The AERONET AOD of level 2.0 screened and quality-assured) is selected to validate the aerosol retrieval results. We use the average AOD of 20×20 pixels around the AERONET site to compare with the AOD acquired from AERONET (Table 2). Table 2 shows that the retrieval AOD results agree well with those of AERONET observations. Among the eight validation cases,the maximum error is 0.051, which is able to meet the AOD needs of atmospheric correction as mentioned in section 2.1.

The reflectance can be retrieved with the following equation.

Based on 6S radiation transfer code, we establish the look-up table (LUT) of coefficients of atmospheric correction in equation (2). The LUT is established as a tree form. To correct image quickly and accurately, we enlarge position and interpolation of the coefficients of atmospheric correction.

Figure 2 is the comparison between images before and after atmospheric correction in the of Poyang Lake and its surrounding areas. Figure 3a is the image before the atmospheric correction and Figure 3b is the image after the atmospheric correc-tion. The image becomes clearer when atmospheric correction has been done. We can see more details such as the river channels from the image when the atmospheric effect is removed.

There are three validation methods which can be used to evaluate the performance of the atmospheric correction.

(1) Comparison of the retrieved reflectance of HJ-1A/1B CCD after atmospheric correction with those measured by Analytical Spectral Devices (ASD) instrument from ground. Note that the ASD can obtain a continuous spectral curve of the reflectance in the visible/near-infrared wavelength, and the reflectance retrieved from HJ-1 A/B CCD data are band-based reflectance covering limited spectral width, the measured spectral reflectance needs to be integrated with the spectral response functions of HJ-1 A/B CCD for each band to obtain the corresponding band reflectance. Two types of land cover (vegetation and soil) are used in the comparison. Figure 2a is the comparison of the vegetation reflectance and Figure 2b is the soil reflectance. The continuous spectrum line is the reflectance measured by ASD instrument, and the discrete dots in four groups are the reflectance extracted from HJ-1A/1B CCD images. In Figure 3, it can be seen that the reflectance of vegetation and soil in the four bands have similar trend with the reflectance of ground measurement. The precision is higher when the reflectance is lower (e.g. the vegetation reflectance in visible bands), and the error increases with the increase of the reflectance (e.g. vegetation reflectance in near-infrared band and the soil reflectance in each band).

(2) Comparison with MODIS surface reflectance products. MOD09 product has 7 different bands, centered at 0.66, 0.87, 0.46, 0.56, 1.24, 1.64, and 2.13μm respectively with a spatial resolution of 500 m (Vermote et al., 1999). This product removed the effects of gaseous and aerosol scattering and absorption as well as adjacency effects caused by variation of land cover. Bidirectional Reflectance Distribution Function (BRDF) and atmosphere coupling effects, and contamination by thin cirrus have been considered in the product of MOD09. The product can reflect the actual reflectivity of the surface more realistically (Liang et al., 2002). Although there are some differences in spectral response function (Figure 4) between HJ-1 A/B CCD and the related bands of MODIS, the reflectance of the typical land types are almost the same, according to the simulation with the land surface spectrum by Sun et al. (2011). We compared the reflectance of the same area at the same time with HJ-1 A/B CCD and MODIS products (Figure 5). To avoid the impact of pixels mismatch, the pure pixels in the center of the uniform area in both two kinds of images were selected and the spatial resolution of HJ-1 A/B CCD was transferred to 500 m. Result shows that the reflectance of HJ-1 A/B CCD is well in accordance with the MODIS products (Table 3).