Determination of land surface reflectance using the AATSR dual-view capability

In this study, a method is presented to retrieve the surface reflectance using the radiances measured at the top of the atmosphere for the two views provided by the Advanced Along-Track Scanning Radiometer (AATSR). In the first step, the aerosol optical depth (AOD) is obtained using the AATSR dual-view algorithm (ADV) by eliminating the effect of the surface on the measured radiances. Hence the AOD is independent of surface properties and can thus be used in the second step to provide the aerosol part of the atmospheric correction which is needed for the surface reflectance retrieval. The method is applied to provide monthly maps of both AOD and surface reflectance at two wavelengths (555 and 659 nm) for the whole year of 2007. The results are validated versus surface reflectance provided by the AERONET-based Surface Reflectance Validation Network (ASRVN). Correlation coefficients are 0.8 and 0.9 for 555 and 659 nm, respectively. The standard deviation is 0.001 for both wavelengths and the absolute error is less than 0.02. Pixel-by-pixel comparison with MODIS (Moderate Resolution Imaging Spectrometer) monthly averaged surface reflectances show a good correlation (0.91 and 0.89 for 555 and 659 nm, respectively) with somewhat higher values (up to 0.05) obtained by ADV over bright surfaces. The difference between the ADVand MODIS-retrieved surface reflectances is smaller than ±0.025 for 68.3 % of the collocated pixels at 555 nm and 79.9 % of the collocated pixels at 659 nm. An application of the results over Australia illustrates the variation in the surface reflectances for different land cover types. The validation and comparison results suggest that the algorithm can be successfully used for both the AATSR and ATSR-2 (which has characteristics similar to AATSR) missions, which together cover a 17-year period of measurements (1995–2012), as well as a prototype for the Sea and Land Surface Temperature Radiometer (SLSTR) planned to be launched in the fall of 2015 onboard the Sentinel-3 satellite.


Introduction
The interest in global satellite observations of land properties for application in Earth system science and global climate research is growing (National Research Council, 2004).Surface albedo, defined as the ratio of upwelling to downwelling radiative flux at the surface (Lucht et al., 2000), is one of the most important variables controlling the surface radiation budget.It has been well recognized that the surface albedo is among the main radiative uncertainties in climate modeling (e.g., Hahmann and Dickinson, 2001;Wang et al., 2007).Snow-free albedo is especially important for land surface models that compute the exchange of energy, water, or carbon for various land use categories (Tasumi et al., 2008;Rechid et al., 2009).Land surface albedo is a key input parameter for land cover classification and is also important for remote sensing of clouds (e.g., Taylor and Stowe, 1984;Coddington et al., 2013;Fricke et al., 2014), aerosols (e.g., Kokhanovsky and de Leeuw, 2009;Seidel et al., 2012) and trace gases (e.g., Wagner et al., 2007).
Surface albedo varies spatially and temporally as a result of both natural processes (e.g., vegetation growth, change in soil moisture content, snow aging) and human activity (e.g., deforestation, agriculture, burning).Important factors are the seasonal phenological stage and precipitation.Fur-Published by Copernicus Publications on behalf of the European Geosciences Union.
L. Sogacheva et al.: Determination of land surface reflectance thermore, the orientation of the surface is important: reflectance might increase for non-horizontal surfaces, such as mountain slopes and high vegetation (e.g., Turner et al., 2008).Three-dimensional surface structure (e.g., segmental high vegetation areas, urban areas) causes shadowing, which is a part of the bidirectional reflectance distribution function (BRDF) effect (van Ginneken et al., 1998;Sailor and Fan, 2002).
The determination of land surface albedo is not straightforward.One option is to assign surface albedo to individual surface and vegetation types and combine these with information on land cover to determine the spatial and temporal distribution of the surface albedo.Alternatively, direct measurements can be done at local sites or information can be retrieved from airborne or satellite data.Each of these methods requires a correction for the effect of atmospheric constituents on the measured reflectance (e.g., Manninen et al., 2012).Another complication is that none of these methods measure albedo but surface reflectance for certain geometries and wavelengths, i.e., the fraction of the incoming solar radiation scattered in a certain direction.Obtaining the albedo requires the integration of reflectance over all sun-view geometries.
In this paper we consider the determination of the surface reflectance using satellite-based radiometer measurements.The reflectance measured with a radiometer at the top of the atmosphere (TOA) consists of solar radiation scattered by both the surface and the atmosphere.Hence, retaining either the atmospheric or the surface contribution to the TOA reflectance requires effective decoupling of these two contributions.Traditional methods for estimating the surface shortwave albedo from satellite data include three steps (Tao, 2012): (1) the satellite observations are converted to surface directional reflectance using atmospheric correction algorithms, (2) surface BRDF models are inverted through the fitting of the surface reflectance composites, (3) the shortwave albedo is calculated from the BRDF through angular and spectral integration.Integrals of BRDF functions result in the so-called black-sky (reflection of direct radiation) and white-sky (reflection of diffuse radiation) albedos that convey important information concerning the inherent properties of surface albedo (Wanner et al., 1997).
During the past several decades, remotely sensed surface albedo and reflectance products have been generated using satellite data.The advantage of the use of satellites as opposed to ground-based or airborne measurements is that satellites can provide global coverage during an extended period of time (decades using the currently available spaceborne instruments).Albedo and reflectance anisotropy products (as given by, for example, BRDF), with temporal frequencies varying from daily to monthly and with spatial resolutions varying from 250 m to 20 km, are derived from sensors on polar-orbiting satellites such as MODIS (Schaaf et al., 2002;Strahler and Muller, 1999), MISR (Lyapustin et al., 2006;Martonchik et al., 1998), POLDER (Bacour and Brèon, 2005;Hautecoeur et al., 2007), MERIS (Guanter et al., 2008), AATSR (Grey and North, 2009;Sayer et al., 2010) and CERES (Rivkin et al., 2006).An overview of the satellites and methods to retrieve global albedo is presented in Schaaf et al. (2008Schaaf et al. ( , 2011)).However, disagreements exist between albedo products from different satellite sensors, due to differences in sensors and observation conditions, and in some cases opposing regional and global long-term trends have been reported (Li, 1996;Zhou et al., 2010;Sayer et al., 2012).
To enable the comparison of the surface reflectance retrieved with different satellites, the BRDF has been introduced as a MODIS product (Schaaf et al., 2002).According to Ju et al. (2010), in order to estimate the BRDF, the operational MODIS albedo and anisotropy algorithm makes use of a kernel-driven, linear model of the bidirectional reflectance factors, which relies on the weighted sum of an isotropic parameter and two functions (or kernels) of viewing and illumination geometry.Radiative transfer models can be used to derive one kernel; the other one is based on surface scattering and geometric casting theory.The kernel weights selected are those that best fit the cloud-cleared, atmospherically corrected surface reflectance available for each location globally over a 16-day period (Lucht et al., 2000).Similar kernel-driven schemes are used to obtain BRDF and albedo information from POLDER (Leroy et al., 1997).The MODIS BRDF product is used in the present work for intercomparison of the AATSR-retrieved surface reflectance.
AATSR and its predecessor ATSR-2 provide two viewsnear nadir and 55 • forward -whose capabilities are used in this paper to determine the land surface reflectance.North et al. (1999) were the first to use ATSR-2 data to determine surface reflectance based on a simple physical model of light scattering for the dual-angular sampling of the instrument.The method is based on the angular constraint, which can be used to separate the surface BRDF from the atmospheric aerosol properties without a priori information on the land surface properties.This model can be used to estimate the degree of atmospheric contamination for a particular set of reflectance measurements and to find the atmospheric parameters which allow retrieval of realistic surface reflectances (Grey and North, 2009).North et al. (1999) report that the corresponding mean absolute error in reflectance estimation, defined for a nadir observation at 555 nm, is less than 0.01.The algorithm was applied to the dual-view AATSR data for a number of sites around the world to test its performance over a range of land covers and aerosol types.Results show good agreement (r 2 = 0.70 for all sites combined) between the AATSR-derived estimates of AOD and sun photometer measurements (Grey et al., 2006b).The retrieval performs best over vegetated land covers for biomass burning aerosol types.
The objective of the current paper is to describe and evaluate a different method for the retrieval of the land surface reflectance which is based on the use of the dual-view ca-pability in order to obtain the AOD nearly independently of the surface reflectance and thus use this value for atmospheric correction in the retrieval of the latter, as described in Sect.3.1.This method is essentially different from that presented by North et al. (2009).
The paper is structured as follows.The AATSR instrument is introduced in Sect. 2. In Sect. 3 the algorithm for the retrieval of AOD and surface reflectance is presented.In addition, the data sets used for validation and comparison are described.Results are presented in Sect. 4 and validated in Sect. 5.As an example, seasonal variations in surface reflectance along a transect over Australia are discussed in Sect.6. Conclusions are presented in Sect.7.

The AATSR instrument
The Advanced Along-Track Scanning Radiometer (AATSR) onboard the ENVISAT satellite (2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012) and its precursor ATSR-2 onboard the ERS-2 (providing level 1 data which are used for aerosol retrieval for the period 1995-2003) are dual-view instruments with across-track conical scanning for both views.One view is near nadir and the other one is at a 55 • forward angle.The time between the two views is 150 s along track.The nominal resolution at nadir is 1 km × 1 km and the swath width is 512 km, which results in global coverage in 5-6 days.AATSR has three wavebands in the visible-near infrared (centered near 555, 659 and 865 nm) and four bands in the infrared (centered near 1610, 3700, 10 850, 12 000 nm).The ADV algorithm uses the 555, 659 and 1610 nm wavebands for the aerosol retrieval over land.
ATSR-2 and AATSR were developed to provide highaccuracy measurements of sea surface temperature for use in studies of global climate change.However, both instruments are also successfully used for the retrieval of aerosol properties in the atmosphere over land and ocean (Veefkind and de Leeuw, 1998;Veefkind et al., 1998;Grey et al., 2006a;Robles-Gonzalez et al., 2000, 2003;Thomas et al., 2009;Sundström et al., 2012;Kolmonen et al., 2013;de Leeuw et al., 2013).

Methods
The AATSR surface reflectance retrieval is based on using independently retrieved AOD as an atmospheric correction of the TOA reflectance measured with AATSR retrieved with the AATSR dual-view retrieval algorithm ADV (see Sect. 3.1 and, for example, Veefkind et al., 1998;Kolmonen et al., 2013, for a description of the most recent ADV version).The basic principle of the aerosol retrieval is to match the AATSR-measured TOA reflectance, in cloud-free conditions, to modeled reflectance at the same wavelengths by minimizing the error function.The modeled reflectance is computed with a radiative transfer model for the transmission of so-lar radiance through the atmosphere which includes a variety of aerosol models.The aerosol model used in the ADV is a mixture of four aerosol components (de Leeuw et al., 2013).
The quality of the AOD retrieved using ADV is similar to that from other AATSR algorithms or to that from MODIS and MISR (de Leeuw et al., 2013).Hence, in view of its measurement with the same instrument, the ADV-retrieved AOD is a good choice for atmospheric correction in the retrieval of land properties using AATSR data.The application of ADV to determine AOD and surface reflectance is described in Sect.3.1.The results are validated using data from the AERONET and ASRVN database, which is described in Sect.3.2.ADV-retrieved surface reflectance is compared with the MODIS albedo/BRDF product, which is described in Sect.3.3.

ADV retrieval algorithm
The TOA reflectance measured by radiometers is the sum of the surface and atmospheric reflectances, and hence the retrieval of the surface reflectance requires an effective decoupling of the surface and atmospheric effects, also referred to as atmospheric correction.Cloud reflectance dominates in the TOA signal, and therefore only cloud-free conditions are considered.Thus strict cloud screening is required.ADV utilizes the semiautomatic algorithm to discriminate between cloudy and cloud-free pixels developed by Koelemeijer et al. (2001).This procedure has been automated by Robles-Gonzalez et al. (2003), who developed a threshold method applied to histograms of reflectances measured in an ATSR-2 scene (see also Curier et al., 2009).Four tests are applied using brightness temperatures in the thermal infrared and reflectances and reflectance ratios in the visible and nearinfrared channels.A pixel is classified as cloud-free only if all tests indicate that no cloud is present.Furthermore, since the retrieval results indicate the possible occurrence of clouds due to imperfect cloud-screening, a post-processing step is applied after AOD retrieval as described in Kolmonen et al. (2013).
The measured TOA reflectance ρ is given by Eq. ( 1) (Chandrasekhar, 1960;Veefkind and de Leeuw, 1998;Kolmonen et al., 2013): where ρ a is the atmospheric reflectance due to aerosol particles and gases (ρ a = ρ aerosol + ρ gas ), ρ s is the surface reflectance, T is the product of downward and upward atmospheric total transmittance, s is the spherical albedo of the atmosphere, and λ is the wavelength.The Sun-satellite geometry is determined by the solar zenith angle µ 1 , the viewing (satellite) zenith angle µ, and the relative azimuth angle between the Sun and the satellite φ.Usually, the surface albedo instead of the surface reflectance is used in the denominator.The choice of using the surface reflectance ρ s together with the spherical albedo s in the denominator of the second term of Eq. ( 1) allows surface reflectance to be solved as described below.The choice made here implies that surface reflectance is assumed to be Lambertian, i.e., surface reflectance is isotropic.However, as applied to anisotropic surface, it is not rigorous.The rigorous analytic solution (not used in ADV) was provided in Lyapustin and Knyazikhin (2001).
The AATSR instrument has two views.In the ADV aerosol retrieval algorithm the surface reflectance is accounted for by using both views and assuming that the ratio of the forward and nadir surface reflectance (the so-called k ratio) is independent of wavelength for the employed AATSR wavelengths (Flowerdew and Haigh, 1995).The k ratio is determined at 1610 nm assuming that the contribution of aerosols and gases to the TOA reflectance is negligible at this wavelength.This assumption does not hold in the presence of large aerosol particles, such as desert dust or sea spray.For other types of aerosol, consisting predominantly of submicron particles, the k ratio can be determined at 1610 nm and used to eliminate surface effects to the TOA reflectance and thus retain the path radiance.The gaseous contribution can be estimated using the atmospheric pressure and temperature, and thus the aerosol contribution is retained.The AOD is retrieved by comparison of the aerosol reflectance with modeled reflectance, determined for a number of aerosol models, each consisting of a mixture of four different aerosol components (de Leeuw et al., 2013).The optimal aerosol component is determined by least-squares fitting for three wavelengths (555, 659 and 1610 nm) simultaneously.The 865 nm wavelength is not used over land as the k ratio assumption does not hold.
The determined AOD, together with the Rayleigh (gaseous) reflectance, can be used to provide atmospheric correction needed for the retrieval of the surface reflectance.It is straightforward to solve the surface reflectance ρ s from Eq. (1): The determined surface reflectance is an indirect but nearly independent retrieval product.The only assumptions used in this procedure are (1) Lambertian surface reflectance and (2) k ratio assumption (the ratio of the surface reflectances in the forward and nadir views are independent of wavelength and the k ratio can be determined at 1610 nm, where the effect of aerosol particles is assumed to be negligible).

ASRVN
Satellite product validation relies on the availability of independent data for the same quantity, usually from ground-based measurements.For the validation of satellite-retrieved aerosol properties, data provided by the ground-based sun photometer network AERONET (Holben et al., 1998) are commonly used.For the validation of satellite-derived surface reflectance the AERONET-based Surface Reflectance Validation Network (ASRVN) database (Wang et al., 2009) has been developed.ASRVN is an operational processing system which uses ancillary AERONET aerosol and water vapor data, while MODIS TOA measurements are used for atmospheric correction (Wang et al., 2009).
The ASRVN products include the bidirectional reflectance factor (BRF, often called surface reflectance), spectral albedo, parameters used in the RossThick-LiSparse (RTLS) BRF model (Lucht et al., 2000; see Sect.3.3 for more details) and a theoretical normalized BRF (NBRF) computed for a standard viewing geometry, VZA = 0 • and SZA = 45 • for MODIS wave bands 1-7 (http://modis.gsfc.nasa.gov/about/specifications.php).For each AERONET site, ASRVN products are stored in a gridded format with a 1 km resolution for an area of 50 km × 50 km.ASRVN is widely used for product validation (e.g., Lyapustin et al., 2007;Wang et al., 2010;Ramon, 2011) and long-term trend and stability studies (Wang et al., 2009).The main sources of errors in the ASRVN algorithm are the residual cloudiness and variation in MODIS pixel size with scan angle, which increases by a factor of 8 from nadir to the edge of scan (Wang et al., 2011).The second is important in regions with high surface heterogeneity.
ASRVN data are available for the period from February 2000 until May 2008.In the current study ASRVN has been used to validate the ADV-retrieved surface reflectance for the whole year of 2007.RTLS BRF model parameters have been used to calculate the ASRVN surface reflectances for the AATSR solar geometry, at the wavelengths of 555 and 659 nm.
To examine the performance of the retrieval for different surface types, ASRVN locations have been subjectively divided (using the AERONET site description and images) into eight groups, according to the land type and industry/population in the surroundings: forest, plane or steppe, desert, coastal site, urban highly populated industrial (ur-ban_hpi), urban, mountain (elevated > 1000 km) and tundra.It is noted that no AATSR/ASRVN collocated pixels over tundra have been found for the 2007 study period.Statistical analysis has been performed to the whole data set and for different surface types.Results are presented in Sect.5.1.

MODIS BRDF product
The MOD43B1 BRDF/Albedo Model Parameters Product (MODIS BRDF/Albedo product, http://modis.gsfc.nasa.gov/data/atbd/atbd_mod09.pdf) supplies the weighting parameters associated with the RTLS BRDF model that best describes the anisotropy for each pixel (Gao et al., 2005).Three parameters -(1) isotropic scattering, (2) radiative transfer type volumetric scattering (from horizontally homogeneous  leaf canopies), and (3) geometric-optical surface scattering (from scenes containing three-dimensional objects) -are provided for all MODIS spectral bands as well as for three broad bands (0.3-0.7, 0.7-5.0 and 0.3-5.0µm).These parameters (e.g., Roujean et al., 1992) can be used to reconstruct the surface anisotropic effects and thus correct directional reflectance to any needed view geometry.
The BRDF kernel fitting method has been validated by comparing ground-based measurements to reflectance remotely retrieved from other satellites.This comparison leads to the conclusion that the difference is small enough (±0.05) for accurate climate modeling (Lucht et al., 2000).The accuracy of the MODIS albedo products using two sets of coincident field measurements -SURFRAD stations and CART/SGP area -has been investigated by Jin et al. (2003).In both networks, the root-mean-square error (RMSE) was less than 0.0177 and a relatively bias of 0.004 was observed for the MODIS albedo products.The reason for the uncertainties in the MODIS spectral surface albedo is the Lamber-tian approximation, which "flattens" the BRDF shape (Lee et al., 1986;Wang et al., 2010).
The MODIS BRDF model also captures the solar zenith angle dependence of the surface albedo as indicated in field measurements.For the broad range of mixed vegetation and structural types, the overall accuracy of the MODIS albedo remains within a ±10 % margin of error for all solar zenith angles (Román et al., 2013).However, the derived surface reflectance is underestimated at high solar or view zenith angles, where BRDF is high, and is overestimated at low zenith angles, where BRDF is low (Liu et al., 2009).

ADV aerosol optical depth
The AOD retrieved using ADV is used as atmospheric correction to obtain the surface reflectance.Therefore the AOD quality is a key factor which determines the quality of the surface reflectance ADV product.   of these data for atmospheric correction in determination of surface reflectance.

ADV surface reflectance
The land surface reflectance has been retrieved from AATSR for the wavelengths of 555 and 659 nm for the whole year of 2007.Examples of the surface reflectance are presented in Fig. 3, for 555 nm, and Fig. 4, for 659 nm, as monthly aggregated maps for March, June, September and December.
Spatial coverage varies from month to month due to the seasonal changes in solar angle and due to the occurrence of snow and ice.AOD cannot, in general, be reliably retrieved with the ADV over surfaces with very high reflectance (see Sect. 3.1), such as over snow and ice, and thus cannot be used for atmospheric correction.
Variations in the land surface reflectance for the same area relate mainly to the seasonality in the vegetation cover and agriculture/forestry activity.
Surface reflectance patterns are similar for both 555 and 659 nm, although obviously spectral differences exist related to the type of land cover.This is illustrated in Fig. 5, where the differences between the land surface reflectances retrieved at 555 and 659 nm are shown for June 2007.For the retrieved areas, the global difference in the surface reflectance retrieved for these wavelengths is about 2 %.The differences in the surface reflectances at 555 and 659 nm are smaller for dark surfaces (forests, cultivated land surfaces) (0-0.02, or 0-2 %) than for bright surfaces such as steppe or mountains (up to 8-10 %).These results agree qualitatively with results presented by Briegleb et al. (1985).
The validation of the surface reflectance results using the ASRVN data is presented in Sect.5.1; their comparison with MODIS data is shown in Sect.5.2.

ADV surface reflectance validation and comparison
Validation of land surface products is important because their accuracy is critical to the scientific community for various applications.The value of the product for science applica-tions and research depends on the accuracy of the data.Thus, validation of the product is needed for quality estimation.Climate modeling requires albedo with an absolute accuracy of ±0.05 according to Henderson- Sellers and Wilson (1983) and of ±0.02 according to Sellers (1995).

ADV-retrieved surface reflectance validation with ASRVN
For the validation of the ADV-retrieved surface reflectance with the ASRVN data, RTLS BRF model parameters have been used to calculate the ASRVN surface reflectances for the AATSR Sun-satellite viewing geometry at wavelengths of 555 and 659 nm, for an area of 50 km × 50 km around each AERONET station.Only the ASRVN data which were obtained within 1 h of the AATSR overpass have been used.ADV-retrieved surface reflectances have been averaged for the same area.Thus, uncertainty related to "point-to-pixel" comparison has been minimized.However, the validation results might still be influenced by uncertainties related to biophysical, spatial, and seasonal signatures and inhomogeneity (Román et al., 2009).Scatterplots of the ADV and ASRVN surface reflectance at both wavelengths are presented in Fig. 6.The statistical metrics for the whole data set (553 collocated data points) for the wavelengths of 555/659 nm are as follows: r = 0.8/0.9,RMSE = 0.02/0.03and slope = 0.91/1.08.The standard deviation (0.001) is the same for both wavelengths.
The collocated data pairs have further been classified according to land cover (see Sect. 3.2).For each subset of land cover data, the statistical metrics for the correlation between ADV and ASRVN reflectances have been computed using linear regression to obtain the standard deviation (σ ), correlation coefficient (r), root mean square error (RMSE), slope and bias (see Table 1).The highest correlation occurs for brighter surfaces, such as steppe (0.90/0.95 for 555/659 nm).The lowest correlation (0.31/0.61) is obtained for coastal sites, where the 50 km × 50 km area may include a mixed ocean-land surface.The standard deviation for each surface type and wavelength is between 0.002 and 0.003.ADV slightly underestimates the reflectance at 555 nm for brighter (mountain, steppe) surfaces.At 659 nm the overestimation is minor (bias = 0; slope = 1.08).Note that validation is limited by the maximum surface reflectance of 0.35 at 555 nm in the ASRVN database.
One of the reasons for the disagreement between ADV and ASRVN-retrieved reflectance is likely that the ASRVN polynomial coefficients used to compute the directional reflectance are derived using the MODIS TOA measurements accumulated for a 16-day interval as the atmospheric correction.Even though the variations in the exact results with aerosol optical depth are small, they affect the retrieval accuracy by a few percent (Lucht et al., 2000).Another reason for the disagreement is that the surface reflectance measured from a satellite will not be purely bidirectional, but will in-  The absolute (U_abs) and relative (U_rel) uncertainties based on the validation have been calculated for each land cover type at both 555 and 659 nm, using The absolute uncertainty for each of the land cover types and for all types together (Table 2) is about 0.02 for surface reflectance at both 555 and 659 nm.The highest relative uncertainty (Table 2) at 555 nm is observed for forest and mountain regions (27.7 and 28.9 %, respectively), and the lowest for steppe (2.1 %).At 659 nm the uncertainty is more evenly distributed for all land types (10-13 %).
We also studied the dependence of uncertainties on aerosol loading.For 555 nm, for lower (< 0.2) and higher (> 0.2) AOD, the uncertainties for all pixels are 12.8 and 12.6 %, respectively.At 659 nm, the uncertainty for low AOD is higher compared to the uncertainty for high-AOD cases (9.5 and −2.1 %, respectively).In Fig. 7 we compare the ADV-and ASRVN-averaged surface reflectance at 555 and 659 nm for each land cover type.Land surface reflectance varies considerably among the surface types.

Comparison of ADV-retrieved surface reflectance with MODIS data
Intercomparison of products from different sensors offers a simple way to evaluate temporal and spatial consistency in addition to the local validation points offered by AS-RVN.For the comparison of the ADV-retrieved surface re-flectance with MODIS products, MODIS reflectances at the AATSR solar zenith angles have been derived from the MODIS albedo using RTLS BRDF model parameters for collocated pixels.This was done only for snowfree pixels selected by using the MODIS product "Percent snow" from the product MCD43C3 (https://lpdaac.usgs.gov/products/modis_products_table/mcd43c3).Monthly aggregated surface reflectance maps for January and June are shown in Figs.ADV and MODIS is very small (0.01).However, there are differences as illustrated in Figs. 10 and 11.Over bright surfaces the surface reflectance at 555 nm retrieved using the ADV is slightly higher than that from MODIS, but 97 % of the pixels agree to within 0.05 and 86 % agree to within 0.025.For 659 nm the differences are slightly larger.For darker surfaces (forest, tundra), the ADV-retrieved surface reflectance is slightly lower (0.01-0.02) than that from MODIS.These differences are similar to those observed in the validation of the ADV-retrieved surface reflectances against the ASRVN data (Sect.5.1); this would indicate imperfections in the ADV retrieval.However, the observed differences could also in part be due to the limitations of the RTLS BRDF model (see discussion in Sects.3.2 and 3.3).
Very high ADV-retrieved surface reflectances (ADV-MODIS > 0.4, less than 0.01 % of total number of pixels retrieved as shown in the histograms in Figs. 10 and 11) occur in coastal regions and in South America, where ADV might have problems with cloud detection.Low ADV-retrieved sur-face reflectances (MODIS-ADV > 0.4, less than 0.001 % of total number of pixels retrieved in winter months; see histograms in Figs. 10 and 11) are located in northern regions with possible snow melt, where the MODIS 16-day aggregated product is indicated to be snow-free, although the actual MODIS surface reflectance is high (0.4-0.8).In that case the problem would not be with ADV but with the MODIS data.Another explanation for MODIS' overestimation in high-latitude regions is that the use of the MODIS product is recommended only for applications with solar zenith angles smaller than 70-75 • (Liu et. al., 2009).The ADV-retrieved surface reflectance may also be low due to effects of cloud shadows, which are not identified and thus not accounted for in the algorithm.
Scatterplots of the ADV-retrieved surface reflectance at 555 nm compared with MODIS data for January and June are shown in Fig. 12.Similar plots for 659 nm are shown in Fig. 13.The number of collocated points, the r value and the regression equation are given at the top of each plot.The  regression equation indicates that the ADV-retrieved surface reflectance is slightly lower than that of MODIS for low surface reflectance (offset = −0.01 in January; offset = −0.02 in July) and somewhat larger for higher surface reflectance.

Surface reflectance spatial and temporal variation: Australia
The effect of different vegetation types for different seasons is illustrated with an example of an AATSR transect over Australia (118 • E, 35 • S; 148 • E, 18 • S; 0.1 • resolution) For this transect, the solar zenith angle changes from ∼ 57 • in the winter to ∼ 33 • in the summer.In spring and fall, the solar zenith angles are ∼ 48 • and ∼ 45 • , respectively.Such differences are not significant with respect to their contribution to seasonal variations in the surface reflectance and are therefore neglected in our study.Other directional effects, which are related to vegetation growth and canopy closure, are not taken into account either but could, of course, influence the temporal variability (Knobelspiesse et al., 2008;Breunig et al., 2011).

Concluding remarks
Land surface reflectance has been retrieved from the AATSR data using an atmospheric correction based on the indepen-dent AOD retrieval product from the AATSR dual-view algorithm (ADV), as described in Sect.3.1.The surface reflectance has been calculated globally with a resolution of 10 km × 10 km for the AATSR wavelengths at 555 nm and 659 for the year 2007.
The validation with the ASRVN network data shows a good agreement with correlation coefficients of 0.8 for 555 nm and 0.9 for 659 nm and standard deviation of 0.001 for both wavelengths.The absolute error for each of the land types and for all types together is about 0.02 for both wavelengths.This value meets the climate modeling requirements indicated by Henderson- Sellers and Wilson (1983) and Sellers (1993).
The spatial variation has been evaluated by comparison with MODIS data.RTLS BRF model parameters have been used to compute the reflectance provided by ASRVN and MODIS to the AATSR Sun-satellite viewing geometry.Pixel-by-pixel comparison with MODIS surface reflectance shows good agreement.In January the difference between the ADV and MODIS surface reflectance at 555 nm is in the range of ±0.05 for 97 % of the pixels and in the range of ±0.025 for 86 % of the pixels.In July, the differences are similar.For 659 nm the agreement is slightly lower (89 and 79 %, respectively).However, for low surface reflectance the ADV-retrieved reflectance tends to be lower than that from either MODIS or ASRVN, while for higher surface reflectance it tends to be higher.One reason might be that the ADV-retrieved AOD tends to be on the low side for high AODs and thus the atmospheric contribution to the TOA reflection is underestimated, leading to overestimation of the surface reflectance.
The ADV surface reflectance might be potentially used as a surface correction for the land temperature retrieval using AATSR (e.g., Prata et al., 1993).Another possible application of the ADV surface reflectance is the surface correction for the AOD retrieval with MEdium Resolution Imaging Spectrometer (MERIS) onboard the same platform as AATSR (ENVISAT) (von Hoyningen-Huene et al., 2011).The ADV surface reflectance retrieved for 555 and 659 nm might also be used for narrow to broadband albedo conversion in the visible part of the spectrum (Liang, 2000;Lucht et al., 2008), which is sensitive to the land surface types (Liang et al., 2005;Dozier et al., 2009).The assumptions made by Briegleb et al. (1985) imply that a representative contribution to the broadband TOA radiance comes from the 555-750 nm spectral interval.
The 17-year data set available from ATSR-2 (1995-2002) and AATSR (2002-2012) provides an excellent opportunity to study long-term surface reflectance variations.The method presented can also be used with the SLSTR (Sea and Land Surface Temperature Radiometer) instrument, which can be considered to be an extended version of AATSR with some extra features, planned to be launched on the Sentinel-3 satellite in 2015.

Figure 1 .
Figure 1.Monthly aggregated AOD at 555 nm retrieved with ADV for March, June, September and December 2007.

Figure 2 .
Figure 2. Validation of the AOD at 555 nm (left) and at 659 nm (right) retrieved from AATSR using ADV against AERONET AOD for year 2007.Colors and symbols relate to different surface types as explained in the legend.

Figure 3 .
Figure 3. Monthly aggregated surface reflectance at 555 nm retrieved from AATSR with ADV for March, June, September and December 2007.

Figure 4 .
Figure 4. Monthly aggregated surface reflectances at 659 nm retrieved from AATSR with ADV for March, June, September and December 2007.

Figure 5 .
Figure 5. Difference between the monthly aggregated ADVretrieved surface reflectances at 659 and 555 nm for June 2007.

Figure 6 .
Figure 6.Scatterplots of ADV-retrieved surface reflectances versus surface reflectances derived from collocated ASRVN albedo matched to the AATSR solar zenith (SZ) angles for wavelengths of 555 nm (left) and 659 nm (right).Colors and symbols relate to different surface types; see legend.

Figure 8 .
Figure 8. ADV (upper panel) and MODIS-derived surface reflectances matching the AATSR viewing geometry (lower panel) for 555 nm for January (left) and July (right).

Figure 9 .
Figure 9. ADV (upper panel) and MODIS-derived surface reflectance matching the AATSR viewing geometry (lower panel) for 659 nm for January (left) and July (right).
8 and 9 for 555 and 659 nm, respectively, for ADV (top) and MODIS (bottom).The surface reflectance patterns retrieved with ADV and MODIS are similar.The averaged global difference between

Figure 10 .
Figure 10.Monthly aggregated maps (upper panel) and histograms (lower panel) of the differences between ADV-retrieved and MODISderived surface reflectances at 555 nm for January (left) and July (right).Numbers in the histogram bins (colored in blue, yellow and red) at the top of the histograms are the percentages of hits of the differences to bins.

Figure 11 .
Figure 11.Monthly aggregated maps (upper panel) and histograms (lower panel) of the differences between ADV-retrieved and MODISderived surface reflectances at 659 nm for January (left) and July (right).Numbers in the histogram bins (colored in blue, yellow and red) at the top of the histograms are the percentages of hits of the differences to bins.

Figure 12 .
Figure 12.ADV vs MODIS point-to-point surface reflectance for 555nm for January (left) and July (right).Color (legend) represents the frequency of the observations.

Figure 13 .
Figure 13.ADV vs MODIS point-to-point surface reflectances for 659 nm for January (left) and July (right).Color (legend) represents the frequency of the observations.

Figure 14 .
Figure 14.ADV surface reflectance for 555 nm (solid lines) and for 659 nm (dashed lines) for winter (June, blue line), spring (September, light-green line), summer (January, green line) and fall (March, red line) along the transect (35 • S, 115 • E-18 • S, 148 • E) over Australia (bottom right image).Vegetation types (http://www.environment.gov.au/node/21580),related to certain areas along the transect, are shown at the top of the figure.

Table 2 .
The absolute (U_abs) and relative (U_rel) uncertainties between ADV and ASRVN surface reflectances at 555/659 nm, calculated for all collocated pixels in different surface type groups.