Assessment of global climate model land surface albedo using MODIS data

Authors


Abstract

[1] Land surface albedo from the Community Land Model is compared to white-sky (diffuse) and black-sky albedo (direct at local solar noon) from MODIS. Generally, comparisons are more favorable in summer than winter, for visible waveband than near-infrared in regions without snow cover, and for black- than white-sky. In regions with extensive snow cover, the model overestimates white- and black-sky albedo by up to 20% absolute. The snow-free visible and near-infrared black-sky albedo is simulated quite well with biases within ±5% over most of the land surface. However, a large negative model bias was found for the Sahara Desert and Arabian Peninsula, particularly in the near-infrared. The poorer simulation of white- compared to black-sky albedo in vegetated areas implies that the model may be overestimating the increase of albedo with solar zenith angle. These results identify several areas that should have priority in further evaluating and improving albedo in the model.

1. Introduction

[2] Surface albedo, the proportion of total incident solar radiation that is reflected from the Earth's surface, determines in large part the amount of energy available to drive turbulent fluxes of heat and moisture. Climate models are sensitive to the specification of land surface albedo [Charney et al., 1977]. Absolute accuracy requirements for datasets suitable for evaluating climate model land albedo range from ±0.05 to ±0.02 [Henderson-Sellers and Wilson, 1983; Sellers et al., 1995]. Albedo data must also be compatible with climate model spatial and temporal resolution and the spectral distribution of radiation.

[3] Satellite remote sensing has long had the potential of meeting these requirements. However, confounding effects such as atmospheric scattering and absorption, anisotropy, inadequate temporal, spatial and spectral sampling, and narrowband to broadband conversions have limited the utility of satellite data. Numerous increasingly sophisticated attempts have been made to address these limitations on a global scale [e.g., Strugnell et al., 2001]. However, a limited number of these methods have resulted in datasets that are adequate for climate model comparisons [Wei et al., 2001].

[4] The MODIS BRDF/Albedo Science Data Product represents the latest attempt at providing a dataset that is suitable for climate model comparisons [Schaaf et al., 2002]. In this paper, we report on a preliminary comparison of climate model albedo with one year of albedo from the MODIS product. Model albedo is computed from the Community Land Model (CLM2) [Bonan et al., 2002], which is the land surface model for the Community Climate System Model (CCSM2) [Blackmon et al., 2001]. We examine white-sky (diffuse) and black-sky (direct at local solar noon) albedo in two broadbands (0.3–0.7 μm (VIS), 0.7–5.0 μm (NIR)). The objective of this paper is not to provide detailed validation of model albedo but to address issues related to making meaningful comparisons and identify priorities for future research.

2. Methods

2.1. MODIS Albedo

[5] The MODIS BRDF/Albedo algorithm relies on a semi-empirical approach to model bi-directional reflectance as a weighted linear sum of an isotropic constant and geometrical-optical and volume scattering kernels that depend only on viewing and illumination geometry but are derived from physical approximations [Schaaf et al., 2002]. The weight of each kernel is derived empirically from MODIS surface reflectance corrected for atmospheric effects and acquired at multiple view and illumination geometries within a 16-day period, which is the two-repeat orbital cycle. The bi-directional reflectances are integrated over the reflected radiation hemisphere to yield black-sky albedo and integrated further over the incident radiation hemisphere to yield white-sky albedo. If the full inversion does not meet certain requirements for sampling and fit, a magnitude inversion is performed in which an archetypal BRDF associated with the appropriate landcover type is adjusted according to the magnitude of the bi-directional reflectances. Spectral albedo in seven MODIS bands is converted to VIS, NIR, and total shortwave albedo using observed spectrums and instrument spectral response functions [Liang et al., 1999].

[6] Here we used the MOD43B3 MODIS BRDF/Albedo product that provides white-sky albedo and black-sky albedo at local solar noon on a 0.5° resolution Climate Modeling Grid [Schaaf et al., 2002]. The product includes quality flags and the fraction of land and snow within each 0.5° cell. The 16-day products have been quality composited to produce monthly products (the higher quality retrieval is selected from the two 16-day periods associated with that month) for January to December 2001 with a gap in June caused by instrument problems. We aggregated the half-degree product to a T42 grid (approximately 2.8° latitude by 2.8° longitude) compatible with our climate model output using area weighting. The quality flags indicate that the percentage of 0.5° cells that qualify as “good quality” is 31–63% depending on month, while the inversion is classified as a “full inversion” for 31–60% of the cells. Since restricting our analysis to only those cells that are of high quality would eliminate too much data, we use all of the data. The median accuracy of the full inversion MODIS albedo product was expected to be 3–11% relative [Lucht, 1998]. Preliminary post-launch validation suggests relative accuracies of 5% (full inversion) to 10% (magnitude inversion) [e.g., Jin et al., 2003]. In addition, there appears to be up to a 10% low bias that only applies to the broadband pure dry snow albedo [Jin et al., 2002]. Broadband values for wet snow and snowy canopies are not affected. Band-specific albedo over all land cover types (including snow) appear accurate as do the broadband values over all land covers.

2.2. Model Albedo

[7] In CLM2, a grid cell is divided into four primary land cover types: glacier, lake, wetland, and vegetation. The vegetated portion of a grid cell is further divided into patches of up to 4 of 16 plant functional types, each with its own leaf and stem area index, and canopy top and bottom heights. The global distribution of plant functional types and their leaf area index are determined from satellite data [Bonan et al., 2002]. Albedo for the vegetated fraction of the grid cell is a blend of snow, soil, and vegetation albedo computed separately for VIS and NIR wavebands and direct and diffuse radiation. Snow albedo is a function of grain size, soot and solar zenith angle. Soil albedo depends on prescribed soil color type and surface soil moisture. Fractional snow cover is used to blend snow and soil albedo. Canopy albedo is based on a two-stream radiative transfer approach [Bonan, 1996]. Vegetation-dependent leaf and stem optical properties are modified by intercepted snow. Exposed leaf and stem area above snow decreases as snow accumulates above lower canopy height. Climatological model albedo was calculated from a 17 year (1984–2000) T42 climate simulation with CLM2 coupled to the Community Atmosphere Model, the atmospheric component of CCSM2, using observed sea surface temperature and sea ice for this period.

[8] Albedo is dependent on the angular distribution of solar radiation. Therefore, it is important to evaluate the direct and diffuse components of model albedo separately. In the model, for example, differences between total shortwave all-sky, white-sky, and black-sky albedo vary by season and latitude (Table 1). Similarly, it is important to separately evaluate the VIS and NIR components of albedo. Vegetation, for example, has a much lower reflectance in the VIS than NIR, while snow reflectance is higher in the VIS than NIR. Furthermore, direct comparisons of model and satellite-derived total shortwave albedo may be problematic because of different spectral distributions of incoming radiation. The MODIS narrowband to broadband conversion coefficients are derived for average atmospheric conditions and indicate that the VIS and NIR albedos are nearly equally weighted to obtain total shortwave albedo [Liang et al., 1999]. The distribution of incoming radiation between VIS and NIR wavebands in the model, on the other hand, depends on the simulated atmospheric conditions (e.g., atmospheric water vapor) and varies with latitude and season (Table 2).

Table 1. Zonal Averages of Differences Between Model All-Sky (ASA), White-Sky (WSA), and Black-Sky Albedo at Local Solar Noon (BSA) (% Absolute)
  30°S30°N
 ASA − WSA1.3−0.42.4
JanASA − BSA1.42.00.0
 WSA − BSA0.12.3−2.3
 ASA − WSA3.11.21.5
JulASA − BSA0.31.40.7
 WSA − BSA−2.80.2−0.8
Table 2. Zonal Averages of the Ratio of Incoming VIS Solar Radiation to Incoming Total Shortwave (VIS + NIR) Radiation for Diffuse or Direct Radiation in the Model
LatitudeJanuaryJuly
DiffuseDirectDiffuseDirect
30°S0.680.450.690.41
0.630.470.700.46
30°N0.700.400.690.46

[9] Two other issues affect the comparisons here. First, the model may have a positive albedo bias in regions with intermittent snow cover because the MODIS product tends toward a snow-free albedo. If the majority of observations within a 16-day period are snow-free, only those observations are used and the data with snow are thrown out. This eliminates the effects of ephemeral snow that can negatively affect the retrieval of the vegetation/soil BRDF. Model albedo is averaged over all days. Second, the MODIS product includes retrievals over shallow inland water and ocean areas (within the limits of 5 km from shoreline or 50 m deep, whichever comes first). A fill value of 4% albedo was used for MODIS pixels exceeding this limit (deep ocean or inland water) when scaling from 1 km to the 0.5° product. Model land albedo includes shallow/deep inland water but not shallow/deep ocean. We used MODIS land fraction to obtain MODIS albedo for the land only. Furthermore, grid cells containing mixtures of land and ocean were eliminated from the comparison to minimize the influence of MODIS shallow ocean albedo on the comparisons. However, the comparisons may still be biased for grid cells that have significant lake fraction in the model.

3. Results and Discussion

[10] Global comparisons of February and July VIS and NIR white- and black-sky albedo are shown in Figure 1. In general, the comparisons are more favorable in summer than winter, for VIS than NIR in regions without snow cover, and for black- than white-sky albedo. In February, the model generally overestimates albedo in North America and Eurasia north of about 55°N as well as in Greenland. These regions generally have extensive snow cover in both the model and MODIS data (Figure 2). Some of the positive bias in regions with significant lake fraction (e.g., the Great Lakes and the Canadian lakes) could be due to the inclusion of lake albedo in the model because CLM2 accounts for lake freezing and snow cover. The MODIS data does not include retrievals over deep lakes in winter because the land/water mask used is independent of season.

Figure 1.

CLM2 - MODIS white- and black-sky albedo (% absolute) in the visible (VIS) and near-infrared (NIR) wavebands for February and July at about 2.8° resolution.

Figure 2.

(a) Snow fraction from MODIS, and (b) snow water equivalent (mm), (c) non-dimensional snow age from CLM2 for February.

[11] Model albedo is lower than observed in regions closer to the southern boundary of the snow line. This occurs primarily in three regions centered roughly at 45°N; north central U.S. extending to south central Canada, Kazakhstan, and eastern Mongolia. In part, this is due to the fact that the snow line in MODIS is further south than the climatological snow line in the model. The model may also be underestimating the fraction of ground covered by snow for shallow snow depths (Z.-L. Yang, personal communication). The NIR appears to be better simulated than VIS in a latitude range extending from the southern snow boundary to about 65°N. The VIS is better simulated than the NIR north of 65°N.

[12] The albedo of snow-covered land in CLM2 is influenced by four primary factors: the spectral properties specified for pure snow, increase of snow albedo for solar zenith angles larger than 60°, the reduction of albedo due to snow aging (grain growth and dirt/soot), and interactions between snow, vegetation, and ground. Insight into the adequacy of CLM2's snow spectral properties can be gained by comparing summer albedo for the pure snow surface (no snow age or solar zenith angle effects in the model) at Summit, Greenland (Table 3). CLM2 VIS black-sky albedo is within 1% absolute of the observations while the NIR is overestimated by 8%. The model's solar zenith angle parameterization increases the black-sky albedo from summer to winter by 1% absolute in the VIS and 6% in the NIR (Table 3). MODIS albedo decreases by 9% in the VIS and 11% in the NIR. MODIS may be seeing the shadows induced by a rough surface at low sun angles while the model parameterization doesn't account for topography effects. There may also be up to a 10% low bias in the MODIS pure dry snow albedo [Jin et al., 2002]. This could account for some of the model's positive bias, at least at the highest latitudes where vegetation is sparse or non-existent. Because positive biases due to snow spectral properties and solar zenith angle are largest in the NIR, one might expect the winter biases to be more severe for the NIR than the VIS. However, south of about 65°N in regions with extensive snow cover, the opposite seems to be the case. Some of the largest differences between the VIS and NIR albedo biases appear to be correlated with regions with high snow age in the model (Figure 2). A non-dimensional snow age of 0.1 results in a 4% relative reduction in VIS albedo and a 10% reduction in NIR albedo.

Table 3. CLM2 and MODIS Black-Sky Albedo (% Absolute) in the VIS and NIR Wavebands Averaged Over Four Grid Cells Centered on Summit, Greenland (72.5°N, 38.5°W)
  CLM2MODIS
FebruaryVIS9586
NIR7146
JulyVIS9495
NIR6557

[13] The model may also be inadequately representing the interactions between snow, vegetation, and ground. One such interaction mentioned previously is the snow cover fraction that could cause a negative model bias. The model may also be underestimating the canopy density as represented by leaf and stem area in regions with a positive model bias. A sparser canopy would result in more snow-covered ground being exposed. The MODIS LAI product could prove useful in evaluating this interaction in the model. More detailed research is also required to evaluate the model snow spectral properties and parameterizations for snow age and solar zenith angle.

[14] In regions that are snow-free year round, the sign of the bias between model and MODIS is fairly consistent between winter and summer (Figure 1). For example, the model underestimates albedo in the Sahara Desert and Arabian Peninsula in winter and summer, particularly in the NIR. This result disagrees with that of Zeng et al. [2002] who found a positive bias in the albedo from the Common Land Model (which uses the same soil albedo parameterization as CLM2) compared to AVHRR-derived albedo. Although the model overestimates the seasonality of precipitation in this region [Bonan et al., 2002], the year-round bias indicates that soil moisture is not the primary cause. In other desert regions of the world (e.g., the Gobi and Australian deserts) the model VIS and NIR snow-free albedo is closer to observed. This indicates that the Sahara Desert and Arabian Peninsula soils have unique spectral properties that should be accounted for in climate models [Tsvetsinskaya et al., 2002].

[15] The VIS black-sky albedo is generally well simulated in snow-free regions in winter and summer. The albedo is within ±5% absolute for most of the land surface and a significant fraction is within ±2%. Some exceptions are southern Africa, parts of Australia, southern South America, and the Canadian Arctic. Negative biases exceeding 5% occur in the Tibetan Plateau and southern Andes and are probably mostly due to differences in snow cover (not shown). Similarly, the NIR black-sky albedo is fairly well simulated although the biases are a bit more pronounced than in the VIS.

[16] The VIS white-sky albedo in summer has much the same biases as the VIS black-sky albedo except the positive biases are somewhat accentuated. The same is generally true for snow-free regions in winter. Similarly, positive biases in the NIR black-sky albedo in summer are accentuated in the white-sky albedo while negative biases are diminished or even become positive. The NIR white-sky albedo is biased quite high over much of the globe by up to 20%. Differences between MODIS VIS white- and black-sky albedo in summer range from less than 1% to 3% and less than 1% to 5% in the NIR (not shown). In contrast, differences between the model VIS white- and black-sky albedo range from less than 1% to 5% and less than 1% to 15% in the NIR. The smallest differences between white- and black-sky albedo in the model are associated with grid cells that have significant areas of bare soil or sparse vegetation (e.g., Sahara Desert and Arabian Peninsula). This is because diffuse soil albedo is assumed to be equal to the direct albedo. The larger differences between white- than black-sky albedo are primarily associated with regions with denser vegetation. Offline simulations with the twostream radiation model indicate that the greatest differences between white- and black-sky albedo occur for plant functional types that have a strong solar zenith angle dependence (dense canopies with random or semi-vertical leaves) [Bonan, 1996]. Further research is needed to explore the relationship between black- and white-sky albedo of vegetation in the model.

[17] The results described here identify several areas that should have priority in further evaluating and improving albedo in the Community Land Model. In particular, more detailed studies are required to examine model parameterizations related to albedo of desert soil and snow-covered land, and the diffuse albedo of vegetation.

Acknowledgments

[18] This work was supported by the NASA Land Cover Land Use Change program through grant W-19,735. The National Center for Atmospheric Research is sponsored by the National Science Foundation. We gratefully acknowledge constructive and helpful comments by two anonymous reviewers.

Ancillary