Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

SCIENCE CHINA

Earth Sciences
• RESEARCH PAPER • April 2015 Vol.58 No.4: 573–588
doi: 10.1007/s11430-014-4997-y

Multi-scale validation strategy for satellite albedo products


and its uncertainty analysis
PENG JingJing1, LIU Qiang2*, WEN JianGuang1, LIU QinHuo1, TANG Yong1,
WANG LiZhao2, DOU BaoCheng1, YOU DongQin1, SUN ChangKui1,
ZHAO XiaoJie1, FENG YouBin2 & SHI Jian1
1
State Key Laboratory of Remote Sensing Science, Institute of Remote Sensing and Digital Earth of
Chinese Academy of Sciences, Beijing 100101, China;
2
College of Global Change and Earth System Science, Beijing Normal University, Beijing 100875, China

Received March 31, 2014; accepted October 8, 2014; published online December 8, 2014

Coarse-resolution satellite albedo products are important for climate change and energy balance research because of their
capability to characterize the spatiotemporal patterns of land surface parameters at both the regional and global scales. The
accuracy of coarse-resolution products is usually assessed via comparison with in situ measurements. The key issue in the
comparison of remote sensing observations with in situ measurements is scaling and uncertainty. This paper presents a
strategy for validating 1-km-resolution remote sensing albedo products using field measurements and high-resolution re-
mote sensing observations. Field measurements were collected to calibrate the high-resolution (30 m) albedo products de-
rived from HJ-1a/b images. Then, the calibrated high-resolution albedo maps were resampled (i.e., upscaled) to assess the
accuracy of the coarse-resolution albedo products. The samples of field measurements and high-resolution pixels are based
on an uncertainty analysis. Two types of coarse-resolution albedo datasets, from global land surface satellite (GLASS) and
moderate-resolution imaging spectroradiometer (MODIS), are validated over the middle reaches of the Heihe River in Chi-
na. The results indicate that the upscaled HJ (Huan Jing means environment in Chinese and this refers to a satellite constel-
lation designed for environment and disaster monitoring by China) albedo, which was calibrated using field measurements,
can provide accurate reference values for validating coarse-resolution satellite albedo products. However, the uncertainties
in the upscaled HJ albedo should be estimated, and pixels with large uncertainties should be excluded from the validation
process.

MCD43B3, GLASS02, uncertainty analysis, validation, remote sensing, albedo

Citation: Peng J J, Liu Q, Wen J G, et al. 2015. Multi-scale validation strategy for satellite albedo products and its uncertainty analysis. Science China: Earth
Sciences, 58: 573–588, doi: 10.1007/s11430-014-4997-y

Over the past two decades, a variety of land-surface albedo al., 2013; Qu et al., 2013) produced in China, and Global
products have been routinely generated from various satel- Land Surface Albedo (GLOBALBEDO) from Satellite Data
lite data, including moderate-resolution imaging spectrora- (Lewis et al., 2012) produced by the European Space
diometer (MODIS) albedo products (Schaaf et al., 2002) Agency (ESA). One way to calibrate/validate (CAL/VAL)
produced by NASA, a global land surface satellite (GLASS) these surface albedo products is to compare reliable surface
albedo product (Liu et al., 2012; Liang et al., 2013; Liu et albedo values taken from field measurements with those
taken from satellite pixels (Liang et al., 2002; Jin, 2003;
Wang et al., 2004; Salomon et al., 2006; Hautecoeur and
*Corresponding author (email: liuqiang@bnu.edu.cn) Roujean, 2007; Susaki et al., 2007; Chen et al., 2008; Liu et

© Science China Press and Springer-Verlag Berlin Heidelberg 2014 earth.scichina.com link.springer.com
574 Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4

al., 2009; Zhang et al., 2010; Román et al., 2011; Cescatti et When both matches are tested and approved, the coarse-
al., 2012; Wang et al., 2012). However, because of the scale resolution product is validated. The advantage of this
discrepancy between point-scale in situ measurements and method is the introduction of the high-resolution albedo
pixel-scale remote sensing observations, the errors caused information. Consequently, the validation results are af-
by different comparison scales greatly affect albedo product fected by the high-resolution albedo uncertainties.
validation. Upscaling field measurements from the point Therefore, although the reference values approximate the
scale to the pixel scale is a challenge in CAL/VAL activities ground truth, the errors in the reference values, which is due
(Cohen and Justice, 1999; Barnsley et al., 2000; Lucht et al., to uncertainties in the validation process and the reference
2000; Liang et al., 2002; Jin, 2003; Susaki et al., 2007; data source, cannot be overlooked. The three main steps of
Roman et al., 2013; Wang et al., 2014). the uncertainty analysis include the following: the identifi-
Three common approaches are taken to reduce scaling cation of all uncertainty sources, uncertainty modeling and
errors. One approach is to set validation sites within homo- uncertainty propagation (Crosetto and Tarantola, 2001).
geneous areas. In this case, surface measurements with These uncertainties come from intrinsic errors within in situ
small instantaneous fields of view (IFOVs) are able to rep- measurements and from external errors caused by naviga-
resent the footprint of coarse-resolution pixels (Liang et al., tion or registration, inaccurate atmospheric correction, and
2002; Wang et al., 2004; Zhang et al., 2010). Previous stud- scale conversion (Liang et al., 2002; Wang et al., 2004;
ies have employed direct “point-to-pixel” comparisons in Buriez et al., 2007; Zhang et al., 2010). The uncertainty
these circumstances (Hautecœur and Leroy, 1998; Barnsley involved in remote sensing products has a significant effect
et al., 2000; Jin, 2003; Salomon et al., 2006; Chen et al., on the accuracy of the modeling and the analysis of
2008; Rutan et al., 2009). Moreover, Susaki et al. (2007) land-surface systems (Li and Wang, 2013; Li, 2014). These
demonstrated that differences in the albedo spatial resolu- uncertainties can result in an overestimation of the error of
tion have negligible effects when the measurement sites can the albedo product (Liang et al., 2002; Wang et al., 2004).
be assumed to be homogeneous. However, completely ho- The absence of uncertainty assessments restricts the appli-
mogeneous areas at the scale of coarse-resolution pixels are cation of multi-scale validation strategies in many ways. For
limited. Another approach is to choose those measurements example, the first match might not be satisfied because of
in which the IFOV is representative of the satellite pixel the limited accuracy of the high-resolution product. With
footprint (Barnsley et al., 2000; Román et al., 2009; Román regard to the second match, the large uncertainty in the ref-
et al., 2011; Cescatti et al., 2012; Wang et al., 2012); these erence value of certain pixels might lead to an exaggerated
measurements are usually gathered from the tops of towers assessment of the satellite product error.
of sufficient height. The representativeness of tower meas- The current study systematically assesses the uncertain-
urements is assessed by analyzing the spatial variation in ties in the reference values for each pixel and designs a
the vicinity of the tower (Lucht et al., 2000; Román et al., multi-scale validation strategy that minimizes uncertainties
2009). Given that only a few field sites qualify, that these and their influence on the validation results. The validation
sites are not evenly distributed on the global scale, and that results are presented with a discussion of the effect of the
they cannot cover all land types, this validation approach is uncertainty control strategy.
limited (Roman et al., 2013).
The third approach, which utilizes high-resolution, re-
motely sensed imagery to upscale the point-scale measure- 1 Methodology
ments to coarse resolutions (Liang et al., 2002; Wang et al.,
2004), is applicable for more heterogeneous areas. The ear- The framework for the multi-scale validation of the coarse-
liest prototype of this method used high-resolution images resolution albedo products is shown in Figure 1. Briefly,
to derive the fractional ground cover of different scene ele- three major steps are involved: (1) collecting the in situ
ments and assigned typical albedos to each scene element. measurements; (2) calibrating the high-resolution albedo
The coarse-resolution albedo is subsequently predicted us- maps with the in situ measurements; and (3) assessing the
ing a linear mixture model. The limitations of this approach uncertainty level of each coarse-resolution pixel using the
include the effects of land-cover classification errors and the co-area high-resolution pixels. Using this framework, sev-
representativeness of the albedo values (Barnsley et al., eral procedures are designed to control the uncertainties in
2000; Lucht et al., 2000). Liang improved the inversion the validation based on the analysis of the uncertainty sources
method for high-resolution albedo and directly used the data and their propagation. For the in situ measurements, a sam-
as the upscaling bridge (Liang et al., 2002). This multi-scale pling scheme is employed to catch the heterogeneity on
validation procedure was summarized by Zhang (Zhang et both the 30 m and the 1 km scales. Each coarse-resolution
al., 2010) as performing “two matches”: first, the in situ pixel is scored according to its total uncertainty resulting
measurements are matched with the high-resolution remote from geometric misalignment and nonlinearity in the albedo
sensing products; then, the high-resolution remote sensing upscaling. Then, the candidate pixels with acceptable un-
products are matched with the coarse-resolution product. certainty are used in the validation of the albedo products.
Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4 575

Figure 1 The proposed multi-validation procedures for the coarse-resolution albedo products. The solid-color boxes identify the steps that might introduce
uncertainties into the validation, and the gradient-filled boxes identify the solutions to the uncertainties that result from the previous step.

1.1 Uncertainties in the validation and their assessment pose, field measurements are referenced to calibrate the bias
in high-resolution albedo products.
1.1.1 Uncertainty sources and propagation Scale effects and geometric mismatches between two
The key issue in validating satellite albedo products is esti- images introduce uncertainty in the upscaling process. As
mating their reference values at the pixel scale. Using the Wu and Li (2009) summarized, scale effects “refer to the
multi-scale validation framework, the reference value is the contrast of information or the different characteristics at
aggregated co-area high-resolution albedo. However, a ref- different scales”, and their primary causes are surface het-
erence value might not always be a satisfactory approxima- erogeneity and the linearity/nonlinearity of the retrieving
tion of the true surface albedo because of the uncertainties algorithms. In this context, the uncertainty caused by the
in the validation procedure (Figure 2). scale effect is the combined result of spatial heterogeneity
Uncertainty originates from both the data source and the and the retrieval model. As previous studies have revealed,
validation process. In terms of data, both field measure- the linearity/non-linearity of the retrieval model is important
ments and high-resolution albedo maps are subject to errors. for scale effects. Albedo is usually retrieved using a linear
The errors within in situ measurements results from imper- model. However, even a linear model can result in a scale
fect instruments, inappropriate operations, and insufficient effect when non-uniform illumination (shading or sheltering)
samples. Conversely, proper instruments, strict operation or multiple scattering (multiple collisions between light rays
protocols, and reasonable sampling schemes reduce field and the in situ surface) is involved because these effects
measurement uncertainties. The errors in the high-resolution change the amount of reflected radiation. Non-uniform il-
albedo products originate from instrumental limitations, lumination incurs a scale effect only if intra-pixel heteroge-
radiometric calibrations, geometric mismatches, atmospheric neity exists, whereas multiple scattering incurs a scale effect
corrections, and retrieval algorithms. The data errors in- even with a uniform surface (Li et al., 2000). Topology is
clude bias and random errors. Field measurement bias can the primary cause for non-uniform illumination and multi-
be reduced by proper instrument adjustment prior to use and ple scattering between 30 m and 1 km (Wen et al., 2009).
careful operation, whereas random errors in field measure- In contrast, the uncertainty from geometric mismatching
ments can be reduced with repeated measurements. Random is independent of retrieval algorithms. Two types of geo-
errors in high-resolution albedo products are reduced during metric mismatching exist: registration error, which results
the upscaling (i.e., averaging) process, whereas bias is pre- from incorrect operations and is beyond the scope of this
served. Therefore, bias in high-resolution albedo products paper, and intrinsic uncertainty, which cannot be avoided.
should be corrected for as much as possible. For this pur- Because the basic unit of a remote sensing image is the pixel,
576 Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4

Figure 2 Illustration of the predominant uncertainty sources and their contribution to the total uncertainty in a reference value. The left boxes list the main
sources of uncertainty in the validation. After the processing steps shown in the intermediate boxes, these uncertainty sources add up to the total uncertainty
shown in the box on the right.

and resampling occurs several times in albedo product gen- (2)–(5) present the detailed calculation of each item:
eration, it is difficult to improve the sub-pixel registration
accuracy of a coarse-resolution pixel (Tan et al., 2006).  grd
2

 grd
2
 bgrd
2
 , (2)
Therefore, this paper considers sub-pixel geometric mis- n
matching to be an intrinsic feature of remote sensing albedo
products. The uncertainty caused by geometric mismatching  high
2
 high
2

 high
2
  , (3)
directly adds to the final total uncertainty, whereas the other n m
uncertainties are increased or decreased after upscaling.
 gm
2

 
2
2
gm
2
gm1  , (4)
1.1.2 Uncertainties in the validation and their assessment n
The aforementioned uncertainties affect the accuracy of the 2
reference values. The validation strategy reduces the uncer-  scale
2
 (t1 h2  t2  ) f topo . (5)
tainties as much as possible and assesses the residual uncer-
tainty. Therefore, the uncertainty propagation is process- In eq. (2),  grd
2
is composed of the bias in the in situ
based instead of model-based. Because the uncertainty sources 2
measurements bgrd and the random noise in the in situ
are independent, summing the squared uncertainties of dif-
ferent sources and mechanisms is a practical method for measurements grd
2
. If the appropriate protocol is followed
obtaining the final uncertainty in the reference values. Eq. 2
to collect the measurements, bgrd is typically small. The
(1) is proposed to conceptually represent the sources and
propagation of uncertainties as well as to estimate the total effect of grd
2
should be weakened due to repeated sam-
uncertainty in a reference value: pling. Moreover, n is the number of samples used in the
calibration of the high-resolution albedo data.
 tot
2
  grd
2
  high
2
+ gm
2
  scale
2
, (1)
In eq. (3),  high
2
primarily originates from the random
where  tot
2
is the uncertainty in the reference albedo value noise in the high-resolution albedo high
2
, which is weak-
for each coarse-resolution pixel calculated from the up- ened by the two-stage upscaling. The question about bias in
scaled high-resolution albedo. The four items on the right the original high-resolution albedo resurfaces due to eq. (3).
side of the equation estimate the uncertainty derived from However, by calibrating the high-resolution albedo with the
the in situ measurements, the high-resolution albedo image, in situ measurements, the bias in the high-resolution albedo
the geometric matching and upscaling, respectively. Eqs. is eliminated. Moreover, the uncertainty in the in situ meas-
Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4 577

urements is transferred into the calibrated high-resolution values in a high-resolution albedo map. The Gaussian and
albedo product. The uncertainty in the high-resolution al- triangular distributions are used to describe the spatial re-
bedo that is calibrated with the in situ measurements is as sponse characteristics of the MODIS reflectance products in
follows: previous studies (Huang et al., 2002; Tan et al., 2006). Un-
der an ideal hypothesis, the spatial response characteristics
 grd
2
  high
2
  gm
2

 c2 high  bgrd


2
2
 , (6) show a uniform rectangular shape (Huang et al., 2002). The
n templates are shifted up, down, left, and right by a half pixel
which consists of the uncertainty derived from the in situ- to simulate geolocation error (Figure 3). Each pair of tem-
measured albedo, the random-noise of the high-resolution plate and position shifts leads to the possible sample value,
albedo and the geometric mismatch uncertainty in the first Ai. In all, 15 samples were collected for each coarse-resolu-
validation match. The importance of these uncertainties tion pixel;  represents the average of the samples. The
decreases as the sample number n increases. standard deviation of these 15 values estimates the uncer-
In eq. (4),  gm
2 tainty caused by geometric mismatching:
2 is the uncertainty that results from the

geometric mismatch between the in situ measurements and 15

the high-resolution pixels. It is determined using the geo-  ( A  )


i
2

metric registration accuracy and the heterogeneity of the  gm1 = i 1


. (7)
15
sampling area. The geometric registration RMSE for the
high-resolution albedo is confined to less than 0.5 pixels in Equation (5) estimates the scaling uncertainty,  scale , in a
the application scenario. By considering the registration semi-empirical way.  h2 denotes the intra-pixel heterogeneity
uncertainty, the in situ measurement area is extended to 2×2
estimated by the standard deviations of the high- resolution
pixels, and the standard deviation is used to evaluate the
heterogeneity within the measurement area, which is ap- albedo within a coarse-resolution pixel;  denotes the
proximately 0.0007–0.01 in the studied application.  gm1 2 average high-resolution albedo within a coarse- resolution
pixel. A topology factor, ftopo, is used to characterize the
is the geometric mismatch between the coarse-resolution roughness of the terrain, which is defined as the average
albedo product and the upscaled, calibrated high-resolution slope angle of the high-resolution pixels within a
albedo. In a remote sensing image prior to geometric regis- low-resolution pixel. The effect of non-uniform illumination
tration, the IFOV of a pixel can be characterized by its center
is expressed as t1 h2 f topo , and the effect of multiple scatter-
location and spatial response function. For example, the
2
pixels in a MODIS level 1B image have triangular spatial ing is expressed as t2  f topo . t1 and t2 are two empirical
response functions, and their center geo-locations are pro-
constants derived via numerical simulation, which can be
vided in the MOD/MYD03 product with an accuracy of
estimated via linear regression with the simulated values of
50 m (1) at the minimum (Wolfe et al., 2002; Tan et al.,
2006). However, for advanced geo-referenced remote sens- scale,  h2 and  from a sufficiently large set of scenarios..
ing products, the geolocation error is enlarged because of The actual DEM for the upstream portion of the Heihe River
the geometric transformation and re-sampling process com- Basin and remote sensing images that were ortho-corrected
bined with the variations in the pixel footprint (Tan et al., and terrain-corrected are adopted as scenarios. By assuming
2006). Here, a pixel in a geo-referenced image is referred to that the surface of each high-resolution pixel is Lambertian,
as a nominal pixel, whose center is located in the regular the incoming illumination, osculation, multiple scattering
grid point; a pixel in the image prior to geometric correction and outgoing radiance in the hemisphere can be simulated
is referred to as an original pixel. The physical parameters based on the method described by (Wen et al., 2009). The
of nominal pixels are derived from the raw pixels whose reflected radiation of the high-resolution pixels accumulates
center locations are not exactly located in the grid point. in the coarse-resolution pixel to derive the actual albedo of
Thus, the actual location of these physical parameters is not the coarse-resolution pixel; the difference with the upscaled
the location of the nominal pixel, and an indelible uncer- albedo in which topographic effects are not considered rep-
tainty exists in the location of the remote sensing product. resents the scaling uncertainty scale. The linear regression
Therefore, half a pixel might be a reasonable estimate of the results that are adopted in this study are t1=0.1108, t2=
geolocation displacement of a coarse-resolution albedo 1.1501.
product. The spatial response function of the nominal pixel
is even more difficult to characterize because the parameter 1.2 Validation strategy
of the nominal pixel sometimes originates from more than
one raw pixel. Here, a statistical method was proposed to 1.2.1 Field sampling scheme
assess the uncertainty due to geometric mismatching: three In the proposed validation framework, the primary goal of
different spatial response functions, Gaussian, triangular, the field campaign is to calibrate the high-resolution albedo
and rectangular distributions, are adopted to sample the map. Morisette et al. (2006) describe the sampling strategy
578 Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4

Figure 3 An illustration of the templates and location shifts used to assess geometric processing uncertainty. The red boxes on the left represent the cov-
ered surface area of a nominal pixel in the 1-km-resolution albedo product, and the black dashed boxes indicate the actual covered area of the physical pa-
rameter.

of VALERI, in which two-stage elementary sampling units 1.2.2 Calibration of the high-resolution albedo
(ESU) were positioned to capture the land-cover character- As Section 2.1.1 explained, the bias in the high-resolution
istics within the research area. The elementary sampling albedo values results from instrument limitations, radio-
unit is referred to as an experiment site, which corresponds metric calibrations, geometric mismatches, and atmospheric
to the coarse-resolution pixel scale. The second-stage sam- corrections. Therefore, it must be calibrated with in situ
pling unit is called a quadrat, which corresponds to the high- measurements. Because the number of field measurements
resolution pixel scale. within a particular site is limited, to increase the stability of
Each experiment site covers at least 2×2 coarse-reso- the calibration, a one-parameter calibration equation is adopted.
lution albedo pixels. This minimum size ensures that the That is, an offset is applied to the entire high-resolution
low-resolution pixels are completely within the experiment albedo map for a particular date, which is estimated using
area given that the maximum geometric displacement is less the average difference between the blue-sky albedo ex-
than 0.5 pixels (Liang et al., 2002). Then, multiple quadrats tracted from the high-resolution albedo map and the in situ
are positioned to capture the variability across each experi- albedo.
ment site. The number of quadrats depends on the site het-
n
erogeneity. It is suggested that at least one quadrat is placed
within each land-cover type in the site (Morisette et al.,  ( qt (i)   hq (i))
C i 1
, (8)
2006). Each quadrat covers at least 2×2 high-resolution n
albedo pixels and multiple measurements are acquired within
a quadrat to capture the variability of the central high-reso- where C is the calibration offset, qt is the averaged quadrat
lution pixel. Because the albedo changes with the solar zen- albedo, hq is the high-resolution albedo that corresponds to
ith angle, the sampling scheme configuration must comprise each quadrat, and n is the number of quadrats within the
between the sampling number and the time cost. site.
Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4 579

1.2.3 Pixel uncertainty rating HJ-1A/B are the first two satellites of the Environment
When validating the coarse-resolution albedo product, it is and Disasters Monitoring Microsatellite Constellation laun-
recommended that only pixels with small geometric mis- ched by China to achieve large-area, all-weather, continuous,
matching, gm1, and scaling uncertainty, scale, be used to and dynamic monitoring of the environment and natural
judge the accuracy of the albedo product. Using the estima- disasters (free download from http://218.247.138.121/
tion method introduced in Section 2.1.2, gm1 and scale are DSSPlatform/index.html). The HJ CCD cameras obtain
calculated for each coarse-resolution pixel. The pixels with 4-band images with a 30 m spatial resolution and a revisit-
a local (3×3 pixel) minimum for gm1 and scale values of less ing circle of less than 2 days. Its high spatial and temporal
than 3% are considered as Class B pixels. Moreover, the resolution makes it an acceptable candidate for the upscal-
pixels for each experiment site are assigned as Class A pix- ing bridge in the validation, helping to minimize the error
els. The remained pixels are classified as Class C pixels. caused by the scale discrepancy. The HJ images were at-
Validation using Class A pixels is similar to the traditional mospherically corrected using the 6S radiative transfer
direct validation method, in which the site average albedo is model with synchronously measured atmosphere parameters
compared with the coarse-resolution albedo product of the that were obtained from a sun-photometer (Sun et al., 2012)
pixel nearest to the center of the site. An alternative option and geo-rectified based on geo-referenced SPOT images
is validation with Class B pixels, in which the validation provided by the Watershed allied telemetry experimental
scope is broadened to an extensive area surrounding the research (WATER) experimental database. However, the HJ
experiment site. Validation with Class C pixels is not rec- CCD sensor has a lower signal-to-noise ratio (SNR) and
ommended due to their significant uncertainty, especially fewer spectral channels than coarse-resolution sensors (e.g.,
when the nonlinearity in the albedo upscaling process has MODIS). Therefore, the accuracy of the HJ albedo map
not been addressed. must be improved by calibration using in situ measurements
before the proper validation of the coarse-resolution albedo
product can be conducted.
2 A case study in the Heihe River area Because the HJ CCD is incapable of providing adequate
multi-angle observations for the albedo retrieval algorithms
2.1 Study area and data description based on the BRDF model inversion, the AB algorithm (the
primary algorithm used to generate the GLASS albedo
2.1.1 Albedo products product) is applied to the HJ albedo retrieval (Sun et al.,
Two global, 1 km albedo products (MCD43B3 and GLASS- 2013). In using the AB algorithm, modifications are made
02A06) are assessed in a validation attempt. The MCD43B3 to adjust the specifications of the HJ-1A/B CCD sensor (Li
albedo product is generated using the MODLAND (MODIS et al., 2000). The details of the AB algorithm can be found
Land) research team of the National Aeronautics and Space in (Qu et al., 2013) The regression coefficients that corre-
Administration (NASA; free download from http://reverb. spond to the surface reflectance values are used in this study
echo.nasa.gov/reverb/). The primary algorithm of this product because the synchronously measured atmosphere parame-
includes an inversion of the Ross-Thick/Li-Sparse-Recip- ters are available and a relatively accurate atmosphere cor-
rocal (RTLSR)-kernel-driven bidirectional reflectance dis- rection is expected. The HJ albedo calibrated using in situ
tribution function (BRDF) model with combined MODIS/ measurements is denoted by the CHJ albedo below.
Terra+Aqua observations over a 16-day window. This algo- To validate the satellite-derive albedo products, including
rithm derives a narrowband albedo via kernel integration the MCD43B3, GLASS02A06 and HJ albedo, the black-sky
and the narrowband-to-broadband conversion (Nightingale albedo and white-sky albedo are combined into the blue-sky
et al., 2008); it has a spatial resolution of 1 km and a tem- albedo using the measured sky-light ratio.
poral resolution of 8 days.
GLASS02A06 (free download from http://glass.bnu. 2.1.2 The in situ albedo
edu.cn/index.html) is the first version of the global 1 km A test of this validation strategy was conducted from July to
albedo product generated by GLASS, a research project August 2011 in the Heihe River Basin (37.7°–42.7°N,
conducted by the National High Technology Research and 97.1°–102.0°E), which is a typical inland river basin in
Development Program of China . The product is derived northwestern China (Li et al., 2013). The landscapes in this
from combined MODIS/Terra+Aqua data at a 1 km spatial area are diverse; they include mountains, forests, steppes,
resolution and an 8-day temporal resolution. The primary oases, and the Gobi desert. Ten experimental sites were
algorithm includes two parts: First, the daily shortwave al- chosen, distributed over the middle and lower reaches of the
bedo is estimated using the angular bin (AB) direct estima- basin. The sites contain various land cover types including
tion algorithm (Qu et al., 2013); then, the daily albedo data desert, oasis, sparse grassland, Gobi, and alkali, and each
are post-processed with a statistics-based temporal filtering one covered an area of at least 3 km×3 km. The sites are
(STF) algorithm to fill gaps and reduce noise (Liu et al., listed in Table 1 and Figure 4. The experiment was con-
2012). ducted for 36 days, including 13 days of measurements (one
580 Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4

Table 1 A list of experimental sites

Station name Station ID Location Land-cover type Measurement date Sampling scheme (Figure 6) Number of quadrats
38.764°N, July 29,
Huazhaizi HZZ_D Sparse grass a 9
100.317°E August 8, 27
38.740°N,
Dangzhai (A) DZ_D Desert July 31 b 6
100.558°E
38.841°N,
Dangzhai (B) DZ_F Farmland July 30 - 6
100.478°E
42.067°N,
Ejin Banner (A) EBA_G Gobi August 2 a 9
100.913°E
42.377°N,
Ejin Banner (B) EBB_G Gobi August 3 a 9
101.148°E
42.209°N,
Ejin Banner (C) EBC_A Alkali August 5 a 9
100.712°E
41.608°N,
Black City BC_G Gobi August 6 b 5
100.920°E
38.925°N,
Baji BJ_D Sparse grass August 9 a 9
100.260°E
38.715°N,
Heping County HPC_D Desert August 25 b 5
100.472°E
38.751°N,
Zhangye Airport AP_D Sparse grass August 22, 28 b 5
100.701°E

Figure 4 The surface albedo measurement sites in the Heihe River Basin. The left map shows the spatial distribution of the experimental sites; the photos
on the right were taken at the experimental sites covering various land-cover types.

or two sites per day); the remaining days had unsuitable site (Figure 6(a)), except for differences in the number of
observation conditions (rain or thick clouds; Figure 5). quadrats due to disparate environmental conditions across
Field measurements were collected in 3 km×3 km quad- several measurement sites (Table 1). For instance, because
rats (Q1–Q9) that were uniformly distributed within each the Dangzhai desert is more homogeneous than the other
Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4 581

Figure 5 The weather record during the experiment.

(a) sampling scheme, one team sampled along the left route,
passing through four quadrats (1-4-7-8, Figure 6(a)), and
one sampled along the right route with the same workload
(3-6-9-5, Figure 6(a)). The observation process usually
lasted 3–4 h at each site, based on the satellite pass time. To
inspect the variance of the albedo during this period, a
CNR4 net radiometer was fixed at a point in Quadrat 2 to
obtain a successive record of the radiant fluxes and albedo
synchronous to the CMP6 measurements. This information
was used to evaluate the effect of the solar zenith angle on
the albedo measurements. The albedo of the experiment
sites in the desert and Gobi areas tended to vary by less than
2% throughout the observation times (usually between 10
am and 3 pm). This variation is one cause of the random
noise within in situ measurements.
The measurement at each point began approximately 20 s
after setup to enable a steady sensor reading. To diminish
possible instrumentation defects, measurements were re-
peated 6 times at each point; then, the two pyranometers
Figure 6 Two types of sampling schemes. placed in a group were exchanged, and another 6 measure-
ments were recorded at the same point. Because of the co-
sine response of the sensor, the height of the albedometer
measurement sites, six quadrats were observed there on July above the surface determines the footprint of the surface
31, 2011. Because of cloudy weather on August 25, 2011, albedo (Cescatti et al., 2012). In our experiment, 90% of the
five quadrats were measured when clouds did not obscure measured radiation originated from a circular area with a
the sun (Figure 6(b)). In addition, the Dangzhai farmland diameter of 7.8 m. The in situ albedo was calculated using
site contained several types of crops; thus, the quadrats were the ratio of upwelling and downwelling radiation. Because
divided by each crop type rather than being uniformly dis- the scale difference between the in situ measurements (ap-
tributed. The area of each sampling quadrat was 60 m×60 m, proximately 7.8 m) and the corresponding HJ pixels (30 m)
which is four times larger than a 30 m×30 m high-resolution was small, the average of the measured albedos at four
pixel. Four sampling points (P1–P4) were set within each sampling points in a quadrat was calculated as the “true
quadrat at intervals of 40 m (Figure 6) to capture the heter- albedo” of the quadrat.
ogeneity within the quadrat.
The upwelling and downwelling radiation fluxes were 2.2 Multi-scale validation results
measured with Kipp and Zonen CMP6 shortwave pyranom-
eters and a CNR4 net radiometer (http://www.kippzonen. 2.2.1 In situ albedo and HJ albedo
com). These instruments were leveled for a minimum view- For the corrected HJ albedo, Figure 7 presents a scatter dia-
ing angle before each measurement. Two CMP6 pyranome- gram of the absolute difference in the pixel-scale HJ albedo
ters were placed in a group to measure upward/downward from the quadrat-measured albedo and the relative deviation
radiation simultaneously. The CNR4 net radiometer ob- (standard deviation divided by the average) of samples
tained both upward and downward radiation measurements. within each quadrat. These quantities are positively corre-
Two teams sampled the quadrats simultaneously. In the lated (R2=0.7005), which indicates that inconsistencies
582 Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4

The corresponding maps of the geometric mismatching


uncertainty,  gm2 , and the scaling uncertainty,  scale , are
illustrated in Figure 10(b) and (c). These maps show that the
pixels with high  gm1 values are primarily distributed
along the borders of different land-cover types and in rug-
ged areas, whereas the pixels with high  scale values are
concentrated in the southwestern mountainous regions.
The temporal correspondence between the CHJ albedo
images and the time window of each coarse-resolution al-
bedo product is presented in Table 2. Eight matched image
groups were found. The summarized validation results for
all 8 groups are displayed in the scatterplots in Figure 11.
The three classes of coarse-resolution pixels are displayed
in different colors. Class A pixels show the highest con-
sistency between the coarse-resolution albedo and the up-
Figure 7 Correlation between the absolute error of the HJ albedo in each
quadrat and the homogeneity of the quadrat for the Huazhaizi experimental scaled CHJ albedo. The error of Class B pixels is slightly
site on August 27, 2011. larger than that of Class A pixels. The root mean square
error (RMSE) is 0.0216 (12.1%) for MCD43, whereas the
between the remote sensing retrievals and in situ measure- RMSE is 0.0185 (10.4%) for GLASS02. The biases for
ments are often associated with strong heterogeneity within MCD43 and GLASS02 are 0.0092 and 0.0127, respectively
the quadrat. (Table 3). Both of these products underestimate the surface
albedo in high-reflectance regions.
2.2.2 HJ albedo and CHJ albedo The RMSE and R2 of each group are displayed in Tables
Figure 8 compares the raw HJ albedo with the in situ albedo 4 and 5. Combined with Figure 11, the error in Class C pix-
for each quadrat. For these comparisons, it was assumed that els is significantly larger than that in Class A and Class B
the surface state would not change within 4 days when no pixels. A large proportion of the Class C points in Figure 11
sudden weather changes occurred. The results show that the clearly deviate from the 1:1 line; this result explains why
HJ albedo generated from the HJ CCD images using the AB different levels of uncertainty affected the reference values
algorithm reflects variations in the in situ albedo. Furthermore, for the three classes. The high uncertainties among Class C
they indicate that a systematic bias exists between the HJ pixels enlarge the final error and lead to an underestimation
albedo and the in situ measurements. The calibration of the of the product’s accuracy. Therefore, the use of all the
HJ albedo based on field measurements can prevent the pro- coarse-resolution pixels in the validation is not suitable.
pagation of this bias during the validation process (Figure 9). Class B pixels are more frequent than Class A pixels, thereby
representing many more land-cover types and comprehen-
2.2.3 Coarse-resolution albedo sively evaluating the quality of the coarse-resolution albedo
For validating the coarse-resolution albedo with the up- products. The validation results for Class B pixels show that
scaled CHJ albedo, a rating of the pixel uncertainty is con- the errors for MCD43B3 and GLASS02A06 are approxi-
ducted to avoid unwanted large uncertainties (Section 1.2.3). mately 0.02 in the middle reaches of Heihe; this value is
The comparison results from August 28, 2011 exemplify the lower than the results from a previous validation study us-
necessity of the pixel uncertainty rating (Figure 10). ing point observations (Tan et al., 2006; Wang et al., 2010)
The validation pixels are expected to be dispersive and but consistent with the validation results based on a scale
cover various land types. Class A pixels (Figure 10(a)) cov- conversion (Lucht et al., 2000; Liang et al., 2002). The cur-
er the major land-cover types within the study area (e.g., rent validation method improves upon the quantity and rep-
desert, oasis, and sparse grass). Class B pixels are uniformly resentativeness of the validation pixels.
distributed in the research area and better represent the
overall conditions of the study area. The consistency be- 2.3 A discussion of uncertainty control in validation
tween Class B pixels and their nearby pixels is high; thus,
they are representative of the region. The pixels in Class B The method described above is an implementation of the
are not distributed in the border region; hence, they avoid multi-scale validation strategy. This method uses two ad-
the disturbing nearby pixels with regard to radiation and vanced procedures: the calibration of a high-resolution al-
geometric registration. In addition, Class B pixels are not bedo map and a pixel uncertainty rating based on an uncer-
distributed in mountainous areas, which avoids the scaling tainty analysis. The benefit of these two procedures is ana-
effects caused by topography. lyzed below.
Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4 583

Figure 8 Comparisons between the HJ albedo and field measurements. Each diagram shows the sampled HJ albedo from one image and the corresponding
in situ measurements collected from one site; the date is noted at the bottom of each diagram. Quadrat 0 was measured using a CNR4, whereas the other
quadrats were measured using two CMP6.

To understand the significance of correcting the HJ al- Therefore, the error is three times greater without calibra-
bedo map using in situ measurements, we recall eq. (6). tion, and the bias uncertainty is significantly reduced using
Here, bgrd is assumed to be zero, and the root of this calibration procedure.
 grd
2
  h2   gm1
2
can be estimated using the RMSE between Figure 12 demonstrates the composition of the uncer-
tainty in the upscaled albedo of the different classes of pix-
the CHJ albedo and the measured albedo of the quadrats
els in the validation results from August 28, 2011. This fig-
(Figure 9(b)). Using the results from August 28, 2011 as an
ure shows that the calibrated uncertainty accounted for a
example,  c high is approximately 5.871×103, but this
large proportion of the total uncertainty. The other uncer-
value would have been 1.765×102 if the step of calibrating tainty elements due to Class A and Class B pixels are obvi-
the HJ albedo with the in situ measurements was skipped. ously smaller than those associated with Class C pixels.
584
Peng J J, et al.
Sci China Earth Sci
April (2015) Vol.58 No.4

Figure 9 Illustration of error reduced by the calibration of HJ albedo. (a) The error between the HJ and surface-measured albedos; (b) the error between the calibrated HJ and surface-measured albedos.
Each CMP6 quadrat was measured with a CMP6 at four uniformly distributed points. The CNR4 quadrat was measured with a CNR4 at the fixed central point. The average error was calculated from all of
the quadrats’ measurements within a particular site. (a) shows that the average error (the dark, filled diamonds) of the HJ albedo is less than or equal to 0.0206. The corrected HJ albedo is closer to the in situ
measurements than the original HJ albedo (b).
Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4 585

Table 2 The temporal correspondence of albedo images

Date of image acquisitiona)


Group
HJ GLASS02A06 MCD43B3
1 207 (July 26) 207 (199–215) 201 (201–216)
2 211 (July 30) 211 (203–219) 209 (209–224)
3 220 (August 8) 219 (211–227) 217 (217–232)
4 220 (August 8) 220 (212–228) 217 (217–232)
5 237 (August 25) 237 (229–245) 233 (233–248)
6 239 (August 27) 238 (230–246) 233 (233–248)
7 239 (August 27) 239 (231–247) 233 (233–248)
8 240 (August 28) 240 (232–248) 233 (233–248)
a) The information in parentheses represents the time range of the orig-
inal reflectance data used to derive the albedo.

Figure 11 Coarse-resolution scatterplots of the 1-km-resolution albedo


products and the upscaled CHJ albedo.

Figure 10 The distribution of Class A and Class B pixels and two main
With regard to Class A pixels, this effect is due to the flat-
kinds of uncertainty. (a) Class A and Class B coarse-resolution pixels dis- ness and uniformity of the selected measurement sites. With
played on a HJ albedo image; (b) the distribution of geometric mismatch- regard to Class B pixels, this effect is attributed to the con-
ing uncertainty  gm 2 , (c) the distribution of scaling uncertainty  scale . trol of uncertainty from the pixel uncertainty rating. Clearly,
586 Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4

Table 3 Summary statistics of the validation for the pixels of different classes

GLASS02A06 MCD43B3
Pixel set Sample number
RMSE R2 Bias RMSE R2 Bias
Class A 64 0.0105 0.9235 0.008 0.0123 0.9294 0.0103
Class B 797 0.0185 0.7774 0.0092 0.0216 0.7582 0.0127
Class C 24675 0.0366 0.5238 0.0143 0.0397 0.4874 0.0181

Table 4 The RMSE between the coarse-resolution albedo and the corresponding upscaled CHJ albedo, calculated for three pixel classes

Class A Class B Class C


Group
GLASS02 MCD43 GLASS02 MCD43 GLASS02 MCD43
1 0.0344 0.04 0.029 0.034 0.0432 0.0451
2 0.0271 0.0319 0.0252 0.029 0.0312 0.0387
3 0.0083 0.0107 0.0097 0.0159 0.0262 0.0357
4 0.0184 0.0217 0.0144 0.0185 0.0203 0.0231
5 0.0103 0.0124 0.0107 0.0163 0.0705 0.0751
6 0.0092 0.0119 0.0098 0.0174 0.0125 0.0196
7 0.0245 0.0281 0.0237 0.0278 0.0336 0.0361
8 0.0084 0.0075 0.0139 0.019 0.017 0.0229

Table 5 The R2 between the coarse-resolution albedo and the corresponding upscaled CHJ albedo, calculated for three pixel classes

Class A Class B Class C


Scene
GLASS02 MCD43 GLASS02 MCD43 GLASS02 MCD43
1 0.962 0.9607 0.9459 0.9764 0.9182 0.9218
2 0.9588 0.9317 0.9317 0.9196 0.9089 0.8628
3 0.9567 0.9883 0.9557 0.9555 0.9266 0.9021
4 0.942 0.8914 0.9029 0.8494 08168 0.8032
5 0.9585 0.9862 0.9526 0.9655 0.8509 0.8054
6 0.9676 0.9556 0.9653 0.9057 0.9124 0.888
7 0.9512 0.9925 0.9633 0.9731 0.9269 0.9109
8 0.9987 0.9981 0.9101 0.9121 0.8819 0.8938

the proportions of  gm1 and  scale are small for both Class
A and B pixels but large for Class C pixels.
The RMSE between the coarse-resolution albedo product
and the reference albedo both contained uncertainties; thus,
the RMSE was denoted the “apparent RMSE”, which usu-
ally overestimates the error in the product. As the validation
result for the Class B pixels shows, the apparent RMSEs for
the MCD43B3 and GLASS02A06 datasets are 0.0216 and
0.0185, respectively, in which the uncertainty of the refer-
ence albedo is as large as 0.0136 (Figure 12). The uncer-
tainty associated with the reference albedo, and not the real
error of the albedo product, dominates the apparent RMSE.
When the Class C pixels are used, the RMSE increased by
0.0181 (0.0397 minus 0.0216) for the MCD43B3 and
0.0181 (0.0366 minus 0.0185) for the GLASS02A06 com-
pared with the Class B pixels. This increase in RMSE is
similar to the increase in uncertainty of 0.0131 (0.0385 mi-
Figure 12 The composition of uncertainty in the upscaled 1 km HJ albe- nus 0.0254). This result shows that evaluating the real error
do. Within the uncertainty of the original HJ albedo, the shaded segment
denotes the proportion that was calibrated via comparison with the in situ of the coarse-resolution albedo products using Class C pix-
measurements; the left section was transferred to the CHJ albedo. els is impossible.
Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4 587

3 Conclusions Cescatti A, Marcolla B, Santhana Vannan S K, et al. 2012. Intercomparison


of MODIS albedo retrievals and in situ measurements across the global
FLUXNET network. Remote Sens Environ, 121: 323–334
This paper quantitatively analyzed the uncertainty in the Chen Y M, Liang S, Wang J, et al. 2008. Validation of MISR land surface
broadband albedo. Int J Remote Sensing, 29: 6971–6983
multi-scale validation procedure for coarse-resolution albe-
Cohen W B, Justice C O. 1999. Validating MODIS terrestrial ecology
do products and proposed a strategy to reduce the uncertain- products: Linking in situ and satellite measurements. Remote Sens En-
ties within in situ measurements and high-resolution albedo. viron, 70: 1–3
To validate this strategy, a preliminary validation test was Crosetto M, Tarantola S. 2001. Uncertainty and sensitivity analysis: Tools
for GIS-based model implementation. Int J Geogr Inf Sci, 15: 415–437
conducted on two types of coarse-resolution albedo prod-
Hautecoeur O, Leroy M M. 1998. Surface bidirectional reflectance distri-
ucts, which was designed under the guidance of uncertainty bution function observed at global scale by POLDER/ADEOS. Ge-
control. ophys Res Lett, 25: 4197–4200
The estimation of reference values is the core of validat- Hautecoeur O, Roujean J L. 2007. Validation of POLDER surface albedo
products based on a review of other satellites and climate databases.
ing coarse-resolution albedo products. Because the refer-
IEEE, 2007: 2844–2847
ence values for albedo are calculated indirectly, they con- Huang C, Townshend J R G, Liang S, et al. 2002. Impact of sensor’s point
tain uncertainties derived from the data source and the vali- spread function on land cover characterization: Assessment and decon-
dation process. Uncertainty in the reference albedo is con- volution. Remote Sens Environ, 80: 203–212
Jin Y. 2003. Consistency of MODIS surface bidirectional reflectance dis-
trolled by choosing proper sampling schemes, calibrating
tribution function and albedo retrievals: 2. Validation. J Geophys Res,
the high-resolution albedo, rating pixels according to their 108: 1–15
uncertainty level, and so on. These procedures minimized Lewis P, Guanter L, Lopez Saldana G, et al. 2012. The ESA GlobAlbedo
the apparent RMSE of the albedo products. Nevertheless, Project: Algorithm. Geoscience and Remote Sensing Symposium
(IGARSS). IEEE Int, 2012: 5745–5748
the apparent RMSE continues to contain some uncertainty
Li X. 2014. Characterization, controlling, and reduction of uncertainties in
from the reference albedo, which future research should the modeling and observation of land-surface systems. Sci China Earth
address. Sci, 57: 80–87
Using a single-point, in situ measurement to validate a Li X, Cheng G, Liu S, et al. 2013. Heihe Watershed Allied Telemetry
Experimental Research (HiWATER): Scientific objectives and experi-
pixel is considered inappropriate by consensus when vali-
mental design. Bull Am Meteorol Soc, 94: 1145–1160
dating coarse-resolution albedo products. The only appro- Li X, Wang J, Strahler A. 2000. Scale effects and scaling-up by geomet-
priate solution to this problem is to use multi-point, in situ ric-optical model. Sci China Ser E-Technolog Sci, 43: 17–22
measurements or a high-resolution image to estimate refer- Li X, Wang Y. 2013. Prospects on future developments of quantitative
remote sensing. J Geogr Sci, 68: 1163–1169
ence values. However, certain problems remain to be solved
Liang S, Fang H, Chen M, et al. 2002. Validating MODIS land surface
with regard to the use of these methods. The current study reflectance and albedo products: Methods and preliminary results. Re-
combined multi-point, in situ measurements with a high- mote Sens Environ, 83: 149–162
resolution image to perform upscaling. A reasonable num- Liang S, Zhao X, Liu S, et al. 2013. A long-term global land surface satel-
lite (GLASS) data-set for environmental studies. Int J Dig Earth,
ber of in situ measurements ensured the reliability of the 6(suppl): 5–33
reference values, and the joining of the high-resolution im- Liu J, Schaaf C, Strahler A, et al. 2009. Validation of Moderate Resolution
age enhanced the efficiency of the validation. Nevertheless, Imaging Spectroradiometer (MODIS) albedo retrieval algorithm: De-
fieldwork is expensive and laborious, which often limits pendence of albedo on solar zenith angle. J Geophys Res, 114: D01106
Liu N, Liu Q, Wang L, et al. 2012. Mapping spatially-temporally continu-
validation and encumbers the reporting of land surface ous shortwave albedo for global land surface from MODIS data. Hydrol
temporal variation. Emerging technologies such as wireless Earth Syst Sci Discuss, 9: 9043–9064
sensor networks enable researchers to automatically and Liu Q, Wang L, Qu Y, et al. 2013. Preliminary evaluation of the long-term
continually obtain multi-point measurements and provide GLASS albedo product. Int J Dig Earth, 6: 69–95
Lucht W, Hyman A H, Strahler A H, et al. 2000. A comparison of satel-
efficient means to acquire more and better in situ data. With lite-derived spectral albedos to ground-based broadband albedo meas-
these new technologies, this multi-scale validation strategy urements modeled to satellite spatial scale for a semidesert landscape.
might become operational in the near future. Remote Sens Environ, 74: 85–98
Morisette J T, Baret F, Privette J L, et al. 2006. Validation of global mod-
erate-resolution LAI products: A framework proposed within the CEOS
This work was sponsored by the National Basic Research Program of land product validation subgroup. Geosci Remote Sens IEEE Trans, 44:
China (Grant No. 2013CB733401), the National High Technology Re- 1804–1817
search and Development Program of China (Grant No. 2012AA12A300) Nightingale J, Nickeson J, Justice C, et al. 2008. Global validation of EOS
and the National Natural Science Foundation of China (Grant Nos. land products, lessons learned and future challenges: A MODIS case
41273168, 41331171). The authors thank the reviewers for their construc- study. In: Proceedings of 33rd International Symposium on Remote
tive comments and suggestions on the paper. Sensing of Environment: Sustaining the Millennium Development Goals,
Stresa, Italy. http://landval. gsfc. nasa. gov/pdf/ ISRSE_Nightingale.
pdf
Barnsley M, Hobson P, Hyman A, et al. 2000. Characterizing the spatial Qu Y, Liu Q, Liang S, et al. 2014. Direct-estimation algorithm for mapping
variability of broadband albedo in a semidesert environment for MODIS daily land-surface broadband albedo from MODIS data. IEEE Trans
validation. Remote Sens Environ, 74: 58–68 Geosci Remote Sens, 52: 907–919
Buriez J C, Parol F, Poussi Z, et al. 2007. An improved derivation of the Román M O, Gatebe C K, Schaaf C B, et al. 2011. Variability in surface
top-of-atmosphere albedo from POLDER/ADEOS-2: 2. Broadband al- BRDF at different spatial scales (30–500 m) over a mixed agricultural
bedo. J Geophys Res, 112: D19201 landscape as retrieved from airborne and satellite spectral measure-
588 Peng J J, et al. Sci China Earth Sci April (2015) Vol.58 No.4

ments. Remote Sens Environ, 115: 2184–2203 Tan B, Woodcock C, Hu J, et al. 2006. The impact of gridding artifacts on
Román M O, Schaaf C B, Woodcock C E, et al. 2009. The MODIS (Col- the local spatial properties of MODIS data: Implications for validation,
lection V005) BRDF/albedo product: Assessment of spatial representa- compositing, and band-to-band registration across resolutions. Remote
tiveness over forested landscapes. Remote Sens Environ, 113: 2476– Sens Environ, 105: 98–114
2498 Wang K, Liang S, Schaaf C L, et al. 2010. Evaluation of moderate resolu-
Roman M O, Gatebe C K, Shuai Y, et al. 2013. Use of in situ and airborne tion imaging spectroradiometer land surface visible and shortwave al-
multiangle data to assess MODIS-and Landsat-based estimates of di- bedo products at FLUXNET sites. J Geophys Res (Atmospheres), 115:
rectional reflectance and albedo. IEEE Trans Geosci Remote Sens, 51: D17107
1393–1404 Wang K, Liu J, Zhou X, et al. 2004. Validation of the MODIS global land
Rutan D, Rose F, Roman M, et al. 2009. Development and assessment of surface albedo product using ground measurements in a semidesert re-
broadband surface albedo from Clouds and the Earth’s radiant energy gion on the Tibetan Plateau. J Geophys Res, 109: D05107
system clouds and radiation swath data product. J Geophys Res, 114: Wang Z, Schaaf C B, Chopping M J, et al. 2012. Evaluation of moder-
1–19 ate-resolution imaging spectroradiometer (MODIS) snow albedo prod-
Salomon J G, Schaaf C B, Strahler A H, et al. 2006. Validation of the uct (MCD43A) over tundra. Remote Sens Environ, 117: 264–280
MODIS bidirectional reflectance distribution function and albedo re- Wang Z, Schaaf C B, Strahler A H, et al. 2014. Evaluation of MODIS
trievals using combined observations from the aqua and terra platforms. albedo product (MCD43A) over grassland, agriculture and forest sur-
IEEE Trans Geosci Remote Sens, 44: 1555–1565 face types during dormant and snow-covered periods. Remote Sens En-
Schaaf C B, Gao F, Strahler A H, et al. 2002. First operational BRDF, viron, 140: 60–77
albedo nadir reflectance products from MODIS. Remote Sens Environ, Wen J, Liu Q, Xiao Q, et al. 2009. Scale effect and scale correction of land-
83: 135–148 surface albedo in rugged terrain. Int J Remote Sens, 30: 5397–5420
Sun C, Liu Q, Wen J. 2013. An algorithm for retrieving land surface albe- Wolfe R E, Nishihama M, Fleig A J, et al. 2002. Achieving sub-pixel geo-
do from HJ-1 CCD data. Remote Sens Land Resour, 25: 58–63 location accuracy in support of MODIS land science. Remote Sens En-
Sun C, Sun L, Ma S, et al. 2012. Atmospheric correction method based on viron, 83: 31–49
HJ-1 CCD data (in Chinese). J Remote Sens, 16: 826–836 Wu H, Li Z L. 2009. Scale issues in remote sensing: A review on analysis,
Susaki J, Yasuoka Y, Kajiwara K, et al. 2007. Validation of MODIS albe- processing and modeling. Sensors, 9: 1768–1793
do products of paddy fields in Japan. IEEE Trans Geosci Remote Sens, Zhang R, Tian J, Li Z, et al. 2010. Principles and methods for the validation
45: 206–217 of quantitative remote sensing products. Sci China Earth Sci, 53: 741–751

You might also like