Tian
et al
., 2013). There are several reasons for neglecting the
case that
BI
performs better than
IB
in the previous studies.
First, they did not distinguish the vegetation growth stages,
i.e., growth versus senescence, which is critical to the per-
formance of
BI
and
IB
. Second, we used vegetated pixels to
assess the performance of
BI
and
IB
, whereas previous studies
compared two strategies for the whole image. Because
BI
and
IB
produced similar results in non-vegetation areas, using the
entire image for accuracy comparison makes the difference
between
BI
and
IB
less significant (Jarihani
et al
., 2014). Fi-
nally, previous comparisons were based on a limited number
of studies without theoretical analysis. Therefore, the results
of previous studies cannot offer a general conclusion, as also
mentioned in Tian
et al
. (2013).
The concerns regarding the approximations or assump-
tions used in this study are the following. First, the
NIR
and
red bands are assumed to have the same weights. Although
the weights of
NIR
and red bands calculated by
STARFM
are not
exactly identical, this approximation is acceptable because
these two bands of the selected similar pixels have similar
characteristics in the spatial, spectral, and temporal dimen-
sions. For the different spectral bands, the spatial distances
between a similar pixel and the central target pixel are the
same. Spectral similarities measure the homogeneity of a
MODIS
pixel, which is reasonably close in different spectral
bands. The temporal distances measure the change between
input and predicted
MODIS
data, which are also close in the
two bands because similar pixels have similar change pat-
terns, i.e., an essential assumption in most spatiotemporal
data fusion methods. Based on the above analysis, the simpli-
fication of the same weights for the
NIR
and red bands may not
lead to large deviations from the original
STARFM
algorithm.
Second, a linear relation is assumed between the increments
of
NIR
and red bands for the vegetation pixels. These assump-
tions might not hold when
NDVI
is saturated; therefore, the
conclusions might be problematic in tropical areas with dense
vegetation throughout all seasons. However, we think that
STARFM
is more often used in temperate areas, where
NDVI
exhibits significant seasonal change and is thus less saturated.
In summary, the above approximations are reasonable for
many applications and the experimental results verify the
reliability of the theoretical analysis.
Although the mathematical proof is demonstrated to be
reasonable by three experiments, some challenges remain.
First, the analysis was conducted for the blending strategy
that only one base pair of Landsat and
MODIS
images are used
as inputs. If more than one pair of Landsat and
MODIS
images
were input, the comparison between
BI
and
IB
becomes much
more complex. Second, the comparison of
BI
and
IB
is based
on vegetated pixels, and the spectra of non-vegetation, such
as soil, are considered stable. However, the spectra of soil are
easily affected by the soil water content, which may cause
some disturbance. Finally, only
STARFM
was discussed herein.
For other blending algorithms, the blending mechanism is
different and the conclusions of this study may be not ap-
plicable.
In conclusion, we suggest that the choice between
BI
or
IB
depends on the vegetation stages of the prediction date and
input Landsat images when we use
STARFM
to generate high-
resolution
NDVI
data. The results of this study can provide
valuable guidance for users to produce better blended
NDVI
data, although the theoretical analysis was based on some
simplifications. In the future, more conditions and spatiotem-
poral data fusion methods will be considered.
Acknowledgments
This work was supported in part by the National Natural
Science Foundation of China (Grant No. 41571406), in part
by the National Key Research and Development Program of
China (Grant No. 2017YFD0300201), and in part by the State
Key Laboratory of Earth Surface Processes and Resource Ecol-
ogy (Grant No. 2015-KF-02).
References
Bencze, M., 2001. Applications of Jensen’s inequality,
Octogon
Mathematical Magazine
, 9:804–824.
Bhandari, B., S. Phinn, and T. Gill, 2012. Preparing Landsat image
time series for monitoring changes in vegetation phenology in
Queensland, Australia,
Remote Sensing
, 4:1856–1886.
Buckdahn, R., J. Ma, and J. Zhang, 2015. Pathwise Taylor expansions
for random fields on multiple dimensional paths,
Stochastic
Processes and Their Applications,
125:2820–2855.
Chen, B., B.Huang, and B. Xu, 2015a. Comparison of Spatiotemporal
Fusion Models: A review,
Remote Sensing
,
7:1798–1835.
Chen, X., D. Yang, J. Chen, X. Cao, 2015b. An improved automated
land cover updating approach by integrating with downscaled
NDVI time series data.
Remote Sensing Letters
, 6: 29–38.
Dong, T., J. Liu, B. Qian, T. Zhao, Q. Jing, X. Geng, J. Wang, T.
Huffman, and J. Shang, 2016. Estimating winter wheat biomass
by assimilating leaf area index derived from fusion of Landsat-8
and MODIS data.
International Journal of Applied Earth
Observation and Geoinformation
, 49: 63–74.
Figure 7. (a)
BI
and
IB
error histogram for vegetation pixels
in Gwydir; and (b)
BI
and
IB
error histogram for non-
vegetation pixels in Gwydir.
Table 2.
RMSE
,
R
,
AD
, and
AAD
of the predicted
NDVI
values of
the vegetation pixels on 25 October 2004 in Gwydir.
Blending Strategy RMSE
R
AD AAD
IB
0.0738 0.3989 0.0396 0.0495
BI
0.1152 0.2935 0.0782 0.0881
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2018
71
