Next the
OLI QA
band is used to estimate the number of
clear pixels in the scene, and the mean brightness tempera-
ture of all clear pixels in the scene is calculated using the
thermal band data. This value is then used in the
LEDAPS
cloud determination algorithm (Schmidt
et al.
, 2013). Since
the spectral test thresholds set in the
LEDAPS
algorithm are
based on
SR
data, they were modified to apply to the
OLI
TOA
reflectance data. The
LEDAPS
algorithms for adjacent cloud
and cloud shadow were also implemented for
OLI
, with the
cloud shadow spectral test thresholds again being modified
to reflect the
OLI
data range and properties. The resultant
masks are then cleaned similarly to the
LEDAPS
masks for
TM
and E
TM
+ data. Water is removed from slopes greater than 2
percent, cloud shadow is removed where the pixel values fail
the darkness tests, and single pixels are removed from cloud,
water, and snow/ice masks using a moving window filter. The
resulting masks are stored in a single 8-bit image mask per
scene alongside the imagery. Histograms of the image masks
are computed and summaries are stored for later processing.
Best Pixel Algorithm and Composite Generation
Once the single-date imagery have been re-projected, clipped
to tile extents, and refined pixel-level masks have been gener-
ated, the data are composited. Because there are generally
several valid pixels to choose from in a given 100-day window,
an algorithm is needed to determine which pixel to choose for
the final composite image. While several existing compositing
methods have been proposed (see above), the methods used in
this effort combine elements of previous algorithms.
LANDFIRE
requires composite images for two target dates per year, while
minimizing phenological changes, which precludes use of
multi-year compositing methods. Additionally, the desire to
ensure the selected pixels are representative of the available
data makes similarity measures a more desired approach over
the median reflectance of several pixels, which can be skewed
by anomalous values that are not removed in the
QA
process.
The similarity method could also be skewed if several unde-
sired pixels are not removed in the
QA
process, but in general,
this should happen less frequently. Prior to choosing the best
pixel for a given tile location, the scene-based metadata files
are scanned, and all products that were processed to Level 1G
(systematic) or those Level 1T products with questionably high
model-fit residuals (see below) are omitted from the processing
list. Next, the image mask histogram files are sorted based on
number of unmasked pixels and proximity to the target date to
create a processing-order list. Scenes that have the lowest per-
centage of masked pixels are given priority and sorted second-
arily by the absolute value of the image date minus the target
date to give higher priority to images closest to the target date.
For each pixel within a tile, the best-pixel algorithm ex-
amines each input scene in the processing-order list, keeping
at most the first five non-masked image pixels, storing all
spectral bands. This
vector of pixels
is then processed with
the algorithm listed in Table 1, utilizing the cosine similarity
function (Sebastiani, 2002; Qian
et al.
, 2004) for each combi-
nation of pixels where there are three or more valid observa-
tions. The cosine similarity represents the cosine of the angle
between the two vectors, with smaller values indicating more
similarity. The cosine similarity function is given in Equa-
tion 3 where
A
= vector A containing the six reflectance band
values of a pixel from one date,
B
= vector B containing the
six reflectance band values of pixel from a second date, and
Sim
(
A
,
B
) = the cosine similarity of the two pixel vectors. The
best pixel is chosen as the pixel with the most similarity to
other pixels that is closest to the target date.
T
able
1. B
est
P
ixel
A
lgorithm
U
sed
for
C
ompositing
based
on
the
N
umber
of
O
bservations
Number of
Observations Algorithm
0
No data present – pixel is masked
1
One observation – pixel is written to the output
2
Two observations; closest to target date is
written to output
3
Most similar of 3 written to output
(3 cosine similarity tests)
4
Most similar of 4 written to output
(6 cosine similarity tests)
5
Most similar of 5 written to output
(10 cosine similarity tests)
Sim
A B
A B
A B
,
.
cos
.
(
)
= −
( )
= − ⋅
1 0
1 0
θ
(3)
If there are less than three valid observations the pixel
acquired closest to the target date is selected. The best pixel
algorithm maintains the 16-bit pixel range as described above,
with the output in GeoTiff format. In addition, a byte-scaled
version in
ERDAS
Imagine
®1
format is created, scaling the 0 to
10,000 integer reflectance mapping to 0 to 400 and clipping
all values above 255 to 255, following Chander
et al.,
(2009)
to use for
MIICA
change detection processing.
The processing flow for generating tiled composites is
illustrated in Figure 3, which shows the individual scenes
within the tile space, the scenes converted to surface reflec-
tance and re-framed within the tile space, and the combina-
tion of overlapping scenes within the tiles to produce the
final composite. Figure 3 also illustrates the process whereby
multiple overlapping scenes are overlaid and the vector of
pixels used to determine which pixel is chosen to populate
the tiled composite.
Modified MIICA Algorithm and Estimation of Severity
The change detection processing uses a variation of the
MIICA
algorithm. Whereas
MIICA
was developed by the
NLCD
project
for detecting change between multi-year epochs, the
LANDFIRE
RSLC
project requires annual disturbance characterization.
The
MIICA
algorithm used by
LANDFIRE
is a subset of the full
MIICA
implementation used by
NLCD
. Following the methods
presented in Jin
et al.,
(2013), the
LANDFIRE MIICA
process
calculates the differenced Normalized Burn Ratio (
dNBR
),
differenced
NDVI
(
dNDVI
), Change Vector, and Relative Change
Vector Maximum indices for seasonally matched, multi-year
image pairs (e.g., 2010 day 175 and 2011 day 175 composites).
A series of conditional statements evaluating combinations
of the four spectral indices against set thresholds, derived
empirically in areas of diverse land cover types, is used to
determine areas of increasing or decreasing biomass.
Examples of the conditional statements are given in Jin
et
al.
(2013). The result is a series of output maps indicating ar-
eas of increasing and decreasing biomass for each combination
of image pairs.
MIICA
outputs were generated separately for
each target date and for each combination of consecutive years
(e.g., 2010 to 2011) and each combination of current year plus
2 years (e.g. 2010 – 2012) for a total of 10 output images. The
various
MIICA
outputs are then combined manually by image
analysts to produce the
LANDFIRE
annual disturbance layers.
In addition to the
dNBR
and
dNDVI
data calculated by
MIICA
,
a Normalized Difference Moisture Index is computed using
the near-infrared and shortwave-infrared bands (
TM
/
ETM
+
1. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.
578
July 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
