Australia, in southern New South Wales (
NSW
, Australia;
145.07°E, 34.00°S). This area is dominated by paddy fields,
with many easily recognizable small patches. Two subset im-
ages (800 × 800 pixels) were acquired on 04 December 2001
and 05 January 2002 and were selected for the experiments
(Figure 2). This period is the crop growth season, which
means that
NDVI
in January is higher than that in December.
Paired Landsat and simulated
MODIS
images on 04 Decem-
ber 2001 as well as simulated
MODIS
image on 05 January 5
2002 were used to estimate the Landsat
NDVI
on 05 January
2002. Vegetation pixels were labeled by the
NDVI
threshold of
0.5 (Figure 3a). Figure 3b and 3c show the
SE
of
IB
and
BI
for
each vegetation pixel. The
BI
error is smaller than the
IB
error
in 75.69 percent of the vegetation pixels. Figure 4a shows
the error distribution of
BI
and
IB
for the vegetation pixels. It
suggests significant underestimation of
NDVI
by both
BI
and
IB
,
whereas the
IB
error is relatively higher. In Table 1, all accu-
racy indices, including the root-mean-square error (
RMSE
), the
correlation coefficient (R), the average difference (
AD
), and the
absolute average difference (
AAD
), suggest that
BI
performs bet-
ter than
IB
in this case. These results are consistent with the
theoretical analysis in the previous Section. For the non-vege-
tation pixels, there is no significant difference between
BI
and
IB
(Figure 4b) because
NDVI
is stable throughout all seasons.
Experiment with the Images Captured in the Vegetation Senescence Period
The second study site was the Lower Gwydir catchment
(hereafter referred to as Gwydir) (Emelyanovaa
et al
., 2013),
in northern New South Wales (149.28°E, 29.09°S), Australia.
This area contains large parcels of croplands and natural
vegetation. Subset images acquired on 22 August 2004 and 25
October 2004 (senescence season of vegetation) were selected
for experimentation (Figure 5). This is the senescence
season
for the crops, which means that
NDVI
in October is lower than
that in August.
Paired Landsat and simulated
MODIS
images on 22 August
2004 and simulated
MODIS
images on 25 October 2004 were
used to test the effect of
BI
and
IB
. Vegetation pixels were
selected based on the
NDVI
threshold of 0.6 (Figure 6a). Figure
6b and 6c show the
SE
of
IB
and
BI
for the vegetation pixels.
The
IB
error is smaller than the
BI
error in 81.33 percent of the
vegetation pixels. Figure 7a shows the error histogram of
BI
and
IB
for the vegetation pixels. In this case, there is a signifi-
cant overestimation of
NDVI
by both
BI
and
IB
, whereas the
BI
error is relatively higher. In Table 2, all accuracy indices
suggest that
IB
significantly outperforms
BI
. This is consistent
with the theoretical analysis in the previous Section. For
the non-vegetation pixels, there is no significant difference
between
BI
and
IB
(Figure 7b).
Experiment Over Homogeneous Region
The third study site is a forest area at 37.70°N and 77.25°W
(Gao
et al
., 2006; Zhu
et al
., 2010). Images acquired on 24
May 2001 and 11 July 2001 were selected for experimenta-
tion. The major land cover in this region is forest (spruce,
pine, and aspen), with fen and patches of sparse soil. The
land cover patches in this area are large enough to find
homogeneous pixels at the
MODIS
resolution. There is little
land cover change during this period, and the forest growth is
rapid in this region. Similarly, subset images with 800 × 800
pixels were selected for the experiments (Figure 8).
Paired Landsat and simulated
MODIS
images on 24 May
2001 and simulated
MODIS
image on 11 July 2001 were used
to predict Landsat
NDVI
on 11 July 2001. There was no need to
select vegetation pixels because most of the pixels represent
vegetation. Water was masked out to prevent disturbance to
the analysis (Figure 9a). Figure 9b and 9c show the
SE
of
IB
and
BI
for the forest area. The
IB
error and
BI
error are compa-
rable in this case. Figure 10 shows the error histograms of
BI
and
IB
for the vegetation pixels. In this case, there is neither
significant overestimation nor underestimation of
NDVI
for ei-
ther
BI
or
IB
. As shown in Table 3,
IB
and
BI
have similar errors
that are much lower than those in
CIA
(Table 1) and Gwydir
(Table 2). This suggests that both
BI
and
IB
yield good blend-
ing results in homogeneous areas.
Discussion and Conclusions
STARFM
provides an effective solution to generate
NDVI
time-
series data at high spatial resolution by blending
MODIS
and
Landsat data. In practice, there are two blending strategies,
“Blend-then-Index” or “Index-then-Blend”. In this study,
theoretical analysis and experiments were conducted to
compare the blending results of
BI
and
IB
. We conclude that
when the
NDVI
values to be predicted are higher than the
input Landsat
NDVI
values,
BI
performs better. In contrast, the
IB
error is smaller when
NDVI
to be predicted is lower than the
input Landsat
NDVI
. When the images are homogeneous, both
BI
and
IB
errors are small.
The difference between
BI
and
IB
is due to the different
second-order terms in the Taylor expansions in our analysis
of error propagation. The second-order terms exist because of
the nonlinearity of
NDVI
. On the other hand, for linear vegeta-
tion indices, such as the difference vegetation index (
DVI
),
BI
and
IB
are expected to produce the similar blending results.
Our conclusions differ from those of previous investiga-
tions that concluded
IB
outperforms
BI
(Jarihani
et al
., 2014;
Figure 4. (a)
BI
and
IB
error histogram for vegetation pixels
in the
CIA
; and (b)
BI
and
IB
error histogram for non-
vegetation pixels in the
CIA
.
Table 1.
RMSE
,
R
,
AD
, and
AAD
of the predicted
NDVI
values of
the vegetation pixels on 05 January 2002 in the
CIA
.
Blending Strategy RMSE
R
AD AAD
IB
0.2542 0.4252 -0.1816 0.1960
BI
0.1353 0.6370 -0.0875 0.1000
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2018
69
