where the
NDVI
curve for the predominant land-cover type
was at 0.10 below the peak
NDVI
value. All scenes that fell
within this date range that had less than 70 percent cloud
cover were selected. This process is illustrated in Figure 2,
which shows the
NDVI
curve for path/row 36/33 for 2010. The
peak
NDVI
value for the predominant land-cover type, which
was deciduous forest, is 0.7. The deciduous forest
NDVI
curve
fell below 0.60 on 17 June 2010 and 21 October 2010. All
L5 scenes with less than 70 percent cloud cover that were
acquired between these dates were downloaded from
GloVis
.
Each L5 scene was reprojected using nearest neighbor
sampling to
GRS
1980,
NAD83
, Albers Conical Equal Area and
snapped to a
CONUS
grid based upon the
NLCD
2001
PTCC
grid.
All scenes were converted to top-of-atmosphere reflectance
(Chander
et al.
, 2009) and then to surface reflectance using
dark-object subtraction (Chavez, 1988). Cloud and shadow
masks were created for each scene using
FMASK
(Zhu and
Woodcock, 2012), which is a stand-alone executable that
produces cloud, shadow, and water raster masks
google.com/p/fmask/
).
FMASK
was run with default parameters.
Model II Regression Mosaic
The model II regression mosaic was produced using the meth-
od described in Beaty
et al.
(2011). The first step in the model
II regression mosaic procedure is to select two adjacent Landsat
scenes with minimal cloud cover. One of these Landsat scenes
is designated as the reference scene and the other one is the
target scene. All pixels in the overlap region of the Landsat
scenes that do not contain clouds, shadows, water, urban, agri-
culture, or anything else that is highly contrasting are identi-
fied as common usable-areas. Using pixels from these areas,
means and standard deviations are calculated and the model II
regression equation (Equation 1) is applied to the target scene.
y y
X
Y
calib
X
Y
X
Y
= ⋅
+ −
⋅
σ
σ
σ
σ
(1)
where
X
–
is the reference scene pixel mean,
σ
X
is the reference
scene pixel standard deviation,
Y
–
is the target scene pixel
mean,
σ
Y
is the target scene pixel standard deviation,
y
is the
target scene pixel value, and
y
calib
is the calibrated target scene
pixel value.
Considerable human interaction is needed to produce
model II regression mosaics. Selection of appropriate Landsat
scenes is done manually as is the delineation of the com-
mon useable-areas. It is not unusual for the end product to be
visually displeasing. When that happens, the entire process is
repeated with different pairs of Landsat images until an accept-
able product has been created. Creating a model II regression
mosaic for a single Landsat path/row takes two to five days.
Maximum NDVI Composite
NDVI
images were generated for every L5 scene, which were
masked using the
FMASK
-generated cloud and shadow mask.
The maximum
NDVI
for the 15 L5 scenes in each path/row was
determined and the pixel with all its bands from the correspond-
ing L5 scene was included in the path/row composite image. For
the path/row overlap areas, the same procedure was done, but
instead of using 15 L5 scenes, the overlapping L5 scenes were
included increasing the number of L5 scenes to 30 or 45.
Median Composite
The univariate median is defined as the middle value in a set
of ordered one-dimensional values. If there is an even number
of values in the set, the median is defined as the mean of the
middle two values. The median is an attractive statistic to use for
compositing because it is the statistic most resistant to outliers.
Prior to calculating the median values, cloud and shadow
masks generated by
FMASK
were applied to each L5 scene.
Masked and image background areas were excluded from
the median calculations. For the median calculations, each
L5 band was considered separately and, thus, the six values
(bands) in a composited pixel could come from different L5
scenes representing different dates. The compositing procedure
returned two images in addition to the median composite. One
reflected the number of input image pixels that were used in
the median calculations. The other held the dates of the images
from which each median value was taken. This image included
12 layers, allowing up to two dates per band for each pixel.
Two dates were involved in the median calculations in situa-
tions where even numbers of input pixels were encountered.
For L5, bands 3 and 4 are used to calculate
NDVI
([band
4 – band 3] / [band 4 + band 3]), bands 4 and 5 are used to
calculate
NDMI
(normalized difference moisture index, [band
4 – band 5] / [band 4 + band 5]), and Tassel Cap (Crist and Ci-
cone, 1984) is a linear combination of bands 1 to5, and 7. For
the median composite, the pixels for these bands are likely
to come from very different dates.
NDVI
,
NDMI
, and Tassel Cap
datasets were created for the 51 path/row median composite
images and for all 765 L5 scenes to examine how the varying
dates of the median composite affected these datasets.
Additionally, median composites of
NDVI
,
NDMI
, and Tassel
Cap were created using the individual L5 images of these de-
rivatives. These median composite
NDVI
,
NDMI
, and Tassel Cap
Figure 2. Example of an NDVI curve obtained from the USGS Global Visualization Viewer (glovis.usgs.gov) for WRS-2 path/row 36/33 for 2010.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2016
201
