PE&RS November 2015 - page 870

values obtained during the same period are also presented in
Figure 9. The statistical analysis results, showing the correla-
tions between wind speed and the differences between the air
temperature and the retrieved IST/MOD29 product, are summa-
rized in Table 7, where no significant correlation is observed.
This implies that the difference between IST and air tempera-
ture was not impacted by wind speed.
Discussion
According to the experiments, an immediate conclusion is
that the proposed method generates a higher accuracy than
the MOD29. A further analysis follows.
Most of the available in situ records of air temperature in
polar areas are from AWSs. In consideration of the cost and the
power consumption, thermistors and passive shields are used
in AWSs, both of which have been shown to be significantly af-
fected by high solar radiation and low wind speeds (Genthon et
al., 2011). According to some researchers (Hudson et al., 2005;
Hall et al., 2008; Genthon et al., 2011; Shuman et al., 2014), this
uncertainty usually occurs when the downward solar irradiance
exceeds 240 Wm-2 and the wind speed is less than 4 ms-1. This
effect cannot be easily corrected since it is shield dependent.
Therefore, only air temperature values acquired when the wind
speed was greater than 4 ms-1 were considered in this research.
The biases in Tables 2, 4, and 6 were calculated using mean
values, which represent the deviation of the MODIS-based IST
from in situ records. The biases of the retrieved ISTs were 0.72
~ 0.92 K less than for the MOD29 product, indicating that the re-
trieved ISTs were accurate. The biases, which are not presented
in the above comparisons, were also calculated for all the data
from Zhongshan Station and the Ross Ice Shelf (−1.46 K for the
retrieved IST and −2.34 K for the MOD29 product). The negative
bias indicates that the MODIS-based IST is lower than the in situ
air temperature. This is mostly due to the difference between
the surface and air temperature, which commonly exists in the
presence of atmospheric temperature inversion (Marks, 2002).
This bias is consistent with the findings of Hall et al. (2008),
where the MODIS-based LSTs were −2 K lower than the AWS-
derived air temperatures. Furthermore, Koenig and Hall (2010)
found a −3 K bias, and Shuman et al. (2014) found a −5 K bias.
However, some instances (see Figures 7 and 9) indicate that
the MODIS-derived IST is slightly higher than the corresponding
air temperature. Referring to Miller’s research (1956), this phe-
nomenon is probably caused by the mixing of warmer air from
aloft during storms, which was also mentioned in Hudson and
Brandt (2005). Furthermore, clouds are another aspect that can
make the surface warmer, in that they have a scattering effect
on radiation transfer in the atmosphere, which can affect the
incoming solar radiation.
Although the difference between the MODIS-derived ISTs and
the observation-based records is clear, the proposed method
presents a better performance than MOD29. This can be attrib-
uted to the application of polynomial fitting in analyzing the
relationship between water vapor and atmospheric transmit-
tance. A similar conclusion was also presented in Ouaidrari’s
research (2002), in which a quadratic split-window equation
provided better accuracy than a linear split-window equation
for AVHRR-derived LST.
A limitation of this method is that its accuracy is affected by
clouds or fog. The accuracy of the IST retrieval method there-
fore relies on the accuracy of the cloud detection result. Even
though there have been effective methods developed for cloud
detection (Liu et al., 2004; Frey et al., 2008), they are extremely
difficult to apply in the polar areas. The similar spectral reflec-
tance between cloud and ice/snow makes it difficult to auto-
matically identify cloud; therefore, it is not easy to determine
whether the weather condition is “cloud free” or “cloudy.” As a
result, some studies have relied on passive-microwave-derived
IST to solve this problem associated with optical data (Cavalieri,
1984; Germain and Cavalieri, 1997; Cavalieri, 1994). However,
the IST retrieval accuracy from passive microwave data is unsta-
ble since the emissivity is variable, depending on the ice/snow
conditions (such as melt or dry) (Mcfarland et al., 1990).
Conclusions
This paper has proposed an approach to retrieve ice surface
temperature (IST) in the Antarctic region, based on a modified
split-window algorithm (SWA ) and atmospheric transmittance
estimation. Through a MODTRAN simulation and regression anal-
ysis, we propose to build the relationship between atmospheric
transmittance and water vapor using a polynomial form, replac-
ing the traditional linear fitting. In this way, more accurate
results were obtained in our experiments. The effectiveness was
quantified by a comparison with the MOD29 product and AWS
data from Zhongshan Station and the Ross Ice Shelf from 2004
to 2013. The results showed that the proposed method can gen-
erate a higher accuracy than the MOD29 product. In addition,
the influence of wind speed on the differences between IST and
air temperature was characterized, with a correlation between
wind speed and the differences observed for Zhongshan Sta-
tion, but no significant correlation found for the Ross Ice Shelf.
Overall, our study is able to provide technical support and a
processing framework for Antarctic surface melt detection.
Acknowledgments
This research was supported by the projects: the National
Natural Science Foundation of China (41206177 and
91338111); the National High Technology Research and De-
velopment Program of China (863 Program) (2012AA12A304,
2013AA12A301); and the Fundamental Research Funds for
the Central Universities (2042014kf0293).
T able 6. The Accuracy of the Retrieved ISTs and the MOD29 Product for the Six AWSs on the Ross Ice Shelf
Data
AWSs
Carolyn Elaine Gill Margaret Schwerdtfeger Vito
Bias (K)
Retrieved IST −2.46 −1.26 −0.98 −1.35 −2.07 −1.70
MOD29 −3.27 −2.27 −1.97 −2.10 −2.89 −2.69
RMSE (K)
Retrieved IST 2.91 2.02 2.01 2.40 2.47 2.44
MOD29 3.59 2.72 2.60 3.17 3.21 3.49
Table 7. Pearson’ s Correlation Coefficients between W ind Speed and the
Differences between the Air Temperature and the Retrieved IST/MOD29
Product , for the Ross Ice Shelf (N = 165)
The differences between the
air temperature and
the retrieved IST
The differences between the
air temperature and the
MOD29
Wind
speed ns ns
ns: no significant correlation.
870 November 2015 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING