row). The smaller the number, the fewer significant differ-
ences between the image types.
For the western zones (16 and 23), the maximum
NDVI
and
the median composites had the highest number of
PTCC
values
that were significantly different from the
PTCC
values for the
model II regression mosaic (1.80 percent and 1.39 percent
of the total pixels, respectively). The
NLCD
2011 shrub/scrub
land-cover class had the most pixels (36 percent to 89 per-
cent) with
PTCC
significant differences. The deciduous forest
land-cover class had the next highest number of pixels (23
percent to 26 percent) with
PTCC
significant differences.
For eastern zone 48, the maximum
NDVI
composite had the
highest number of
PTCC
values with significant differences
from the
PTCC
values for the median composite and model II
regression mosaic (10.80 percent and 9.24 percent of the total
pixels, respectively). For the other eastern zones (54 and 59),
the maximum
NDVI
composite and the model II regression
mosaic had the highest number of
PTCC
pixels with significant
differences as compared to the
PTCC
values for the median
composite (3.64 percent to 7.28 percent of the total pixels). For
all eastern zones, most of these
PTCC
differences occurred in
the deciduous forest (28 percent to 57 percent of the significant
difference pixels) and pasture/hay (13 percent to 27 percent of
the significant difference pixels)
NLCD
2011 land-cover classes.
Discussion
Studies have shown the efficacy of Landsat composites.
Hansen
et al
. (2008) created Landsat composites by selecting
pixels with the lowest cloud and shadow likelihood values
for mapping forest cover and clearing in the Congo Basin. Roy
et al
. (2010) created monthly, seasonal, and annual Landsat
composites by using maximum
NDVI
and maximum brightness
temperature. Flood (2013) created Landsat composites using
a multi-dimensional median approach. Griffiths
et al
. (2013)
developed composites using a parametric weighting scheme,
which included acquisition dates and distance to clouds.
This study is the first study to create L5 univariate me-
dian composite images to be used to model
PTCC
. The main
concern with using a median composite image for modeling
vegetation is that the integrity of the bands for individual
pixels is not preserved. The analysis showed that despite dif-
ferent dates being used for the bands of individual pixels for
the median composite (Tables 1, 2, and 3), the
NDVI
,
NDMI
, and
Tassel Cap curves resembled the curves for the
NDVI
,
NDMI
,
and Tassel Cap derived from individual L5 scenes. Addition-
ally, the
NDVI
,
NDMI
, and Tassel Cap values derived from the
median composite were strongly correlated with the median
values of these same image derivatives for the individual L5
scenes (Figure 3). The loss of band integrity did not adversely
affect these L5 image derivatives.
The
NDVI
,
NDMI
, and Tassel Cap values derived using the
maximum
NDVI
composite, which does preserve the integrity
of the bands, were similar to the same image derivative values
derived using the median composite (Figure 4). However,
NDVI
,
NDMI
, and Tassel Cap values derived from the model II
regression mosaic were lower than these same image deriva-
tive values derived from both image composites except for
zone 23. For zone 23, all image derivatives were similar. The
reason for the differences in image derivative values between
the model II regression mosaic and the other image compos-
ites is the difference in the way the L5 scenes were selected
for the different procedures. L5 scenes selected for the model
II regression mosaic procedure were chosen based upon image
quality while the L5 scenes chosen for the median and maxi-
mum
NDVI
composites were chosen based upon
NDVI
values.
For the western zones (16 and 23), between 0.002 percent
and 1.8 percent of the
PTCC
values derived using the three
composite types had significant differences (Table 8). Most of
the differences occurred in the
NLCD
2011 shrub/scrub land-
cover class. For the shrub/scrub land-cover class, the
PTCC
values that were derived using the model II regression mosaic
were lower than those derived using the maximum
NDVI
and
median composites. For zone 23, these results support the
Wilcoxon signed rank tests which showed the
PTCC
values
derived using the model II regression mosaic were 2 percent
lower as compared to the
PTCC
values derived using the other
two composite types (Table 7).
The L5 scenes used for zone 16 and 23 to create the model
II regression mosaic were from mid-August to late October
(Table 5). The L5 scenes used for the two composites were
from June to October (Tables 1, 2, and 4), which encompassed
more of the growing season for shrub/scrub. The green-up of
the shrub/scrub resulted in higher
PTCC
values for the maxi-
mum
NDVI
and median composites than for the model II re-
gression mosaic. Since the
PTCC
values were for trees and not
shrubs, the fact that significant differences occurred in shrub/
scrub land-cover types is less important.
For the eastern zones (48, 54, and 59), there were more
significant differences for the
PTCC
values derived using the
different image types, but the percentage of pixels with signif-
icant differences was still low, between 10.8 percent and 0.01
percent (Table 8). The
PTCC
values derived using the maxi-
mum
NDVI
composite and model II regression mosaic had the
largest significant differences as compared to the median com-
posite. There were also high significant differences between
the
PTCC
values derived using the maximum
NDVI
composite
and those derived using the model II regression mosaic. The
majority of these significant differences occurred in the de-
ciduous forest and hay/pasture
NLCD
2011 land-cover classes.
For the deciduous forest land-cover class, 91 to 96 percent of
the pixels with significant differences had lower
PTCC
values
(17 to 18 percent lower) for the maximum
NDVI
composite
and the model II regression mosaic as compared to the
PTCC
values for the median composite. For zone 54, these results
support the Wilcoxon sign rank tests, which showed the
PTCC
values derived using the median composite were significantly
higher than the
PTCC
values derived using the maximum
NDVI
composite (Table 7). The Wilcoxon signed rank tests found no
other significant differences. For the eastern zones, the major-
ity of the pixels (60 percent) for the maximum
NDVI
composite
came from May and June dates while the median composite
pixels came from variable dates and covered the entire grow-
ing season (May to Oct) (Tables 1, 2, and 4). The dates used
for the model II regression mosaic pixels were all from Aug to
Oct (Table 5). The greater diversity of dates and, thus, spectral
values in the median composite led to higher estimates of
PTCC
especially for deciduous forest land cover.
Conclusions
This study showed that univariate median Landsat-5 compos-
ites can be used to effectively model
PTCC
. This conclusion is
based upon three findings.
1. First, regarding the median and maximum
NDVI
com-
posite, there was no loss of information. Even though
the spectral values for the median composite came
from a variety of dates (Tables 1 and 2), image deriva-
tives such as
NDVI
,
NDMI
, and Tassel Cap created from
the median composite and those created from original
Landsat-5 scenes were highly correlated (r
2
>0.97; Fig-
ure 3). The dates used for the median composite came
from a greater variety of dates than the maximum
NDVI
composite (Table 4), but the image derivatives created
from these images were very similar (Figure 4). How-
ever, the image derivatives created from the model II
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2016
209
