The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
96
The AERONET ground-based sunphotometer makes one
measurement every 15 minutes. AOT of 7 channels (0.34, 0.38,
0.44, 0.50, 0.67, 0.87 and 1.02pm) can be derived from the
direct solar radiation measurements (with only part channels at
some sites) with the precision of 0.01-0.02 (Eck et al, 1999).
The ground-based and satellite products used covered the
period from Jan. 2005 to May 2007.
2.2 Spatio-temporal Matching Up Method
The spatio-temporal sampling properties of satellite and
ground-based data are different. Thus, proper sampling window
should be determined according to the measurement interval of
ground-based data, space resolution of satellite product and
aerosol variation and uniformity over time and space. The
averages of effective aerosol pixels in the space window and
effective ground-based measurements in the time window make
one match-up to be involved into the validation dataset.
Based on former studies (Choku et al, 2002; Remer et al, 2002,
2005; Chen & Yang, 2005), ground-based data were averaged
within ±0.5h (or ±1.0h) centered at the satellite observation
time, and those with measurements not less than 2 (or 4) were
included into the validation dataset.
Satellite data were averaged within a radius of 25km (or 45km)
centered at the AERONET site location, and those with
effective pixels not less than 5 (or 9) after data elimination for
land, turbid water, and cloud contamination were included into
the validation dataset.
2.3 Validation Channel Selection
Although MODIS and AERONET provide AOT for several
channels, only channels centered at 658nm (675nm for
AERONET) and 860nm (870nm for AERONET) are most
nearest for direct comparison.
Former research indicated that the natural logarithm of AOT
can be well fitted as the quadratic polynomial of natural
logarithm of wavelength, with the fitting error of about
0.01-0.02 (Eck et al, 1999). Thus, the 550nm AOT at
AERONET site was estimated with equation (1) to validate the
MODIS AOT at 550nm. Wavelengths of 440, 675, 870 and
1020nm were selected for fitting.
lnx a (X,)=ao+ a i lnX,+a 2 (lnX.) 2 ( 1 )
where x a (À) = AOT at wavelength X
ao, a b a 2 = polynomial coefficients
3. RESULTS AND DISCUSSION
Using 25km radius and ±0.5h as the sampling window, there
were 162 match-ups. Excluding those with ground-based
measurements less than 2 and satellite pixels less than 5, 32
matched data were left, in which ground-based data were from
sites of “Gosan SNU” (29 match-ups) and “Shirahama” (3
match-ups) with the averaged AOT at 550nm of 0.35. Figure 2
displays scatter diagrams of the match-ups, in which dashed
lines represent the upper and lower limits of expected
uncertainties of MODIS AOT product (over ocean,
At=±0.05±0.05t for 550nm (Tanre et al, 1997) and
At=±0.03±0.05t (Remer et al, 2002) for 660 and 860nm), and
the red is the linear fitting line. Table 1 lists the corresponding
error statistics.
The R 2 of linear fitting between MODIS and AERONET was
0.68, 0.65 and 0.56 for MODIS 550, 658 and 860nm channels
respectively. There was relatively large dispersion at low AOT,
and the R 2 of these 3 channels was increased to 0.77, 0.74 and
0.67 when data with AOT less than 0.05 were discarded.
The RMSE was 0.15, 0.13 and 0.10 for MODIS 550, 658 and
860nm channels respectively. The mean relative errors all
exceeded 70%, but were reduced to 37.8%, 30.8% and 30.0%
without AOT less than 0.05.
The percentages of retrievals within the expected uncertainties
were 62.5% for 550nm channel (only 46.9% for
At=±0.03±0.05t) and 46.9% for channels of 658 and 860nm.
There were relatively large disparities compared with the
results of Remer et al.(2005) using MODIS aerosol products
from Aug. 2000 to Aug. 2002 and AERONET data from 132
sites globally distributed ( level 1.5 product, 2052 match-ups
with averaged AERONET 550nm AOT of 0.18): 62%, 66% and
70% respectively for 550, 660 and 870nm3 *. But, in that
research, the percentages were only 56%, 53% and 60% over
Asia-Pacific waters (57 match-ups with averaged AERONET
550nm AOT of 0.21). It indicates that the retrieval performance
is relatively poor in this region compared with that of global.
When data with AOT less than 0.05 were discarded, the
percentages were increased to 66.7% for 550nm (50.0% for
At=±0.03±0.05t) and 51.7% for 658 and 860nm channels. If
the uncertainties were At=±0.05±0.05t for 658 and 860nm
channels, the percentages were increased to 59.4% and 65.6%
separately.
Figure 2. AOT match-up scatter diagrams at 550nm, 658nm and
860nm for 25km radius and ±0.5h
The result of “Shirahama” was better with RMSE of 0.03, 0.04
and 0.04, and mean relative error of 8.0%, 13.2% and 20.9%
respectively for 550, 658 and 860nm channels compared with
others. But, it lacked statistical meaning because of the little
data amount. For spring (Mar. to May, 22 match-ups), the linear
*At=±0.03±0.05t, sampling window was 5x5 pixels and ±0.5h,
same data elimination method as in this paper.