PE&RS July 2015 - page 580

specification. However, these scenes were, in some cases,
shifted hundreds to thousands of meters. The issue was traced
to questionably high model-fit residuals and error checks in
the LPGS that were not handling this situation correctly. The
affected scenes were excluded from LF2012 processing, and
the LPGS issue was eventually corrected by USGS.
Masking Results
The LEDAPS-derived data masks that were enhanced and used
for TM and ETM+ performed sufficiently well in most cases,
though some undesired data still existed in the final composites.
Undesired data occurred most often in areas that are con-
tinuously cloudy and was caused by optically thin clouds or
cloud edges that were not captured in the cloud masks. Limit-
ed pixel choices leading to the selection of the closest pixel to
the target date without regard to similarity further exacerbated
this issue. Other cases could exist where three or more pixels
were available to choose from but more than one was cloudy,
and therefore the most similar pixels were cloudy, though
still chosen as the most suitable. Cloud shadows were also
problematic in many scenes. The LEDAPS algorithm uses an
estimate of cloud height based on the brightness temperature
and lapse rate calculations from the surface to then project
cloud shadow positions based on solar position and illumina-
tion geometry. This method works well for some cloud types,
such as mid-level optically thick clouds, though tends to pro-
duce unrealistic shadows for high-level optically thin clouds,
haze, and some contrails. The shadow enhancement methods
described above work well for removing unrealistic shadows
from the mask, but there is currently no mechanism to add to
the mask where cloud shadows are not recognized. This leads
to unmasked cloud shadows being chosen for the composite
in cases where the pixel similarity tests are not performed or
multiple shadowed pixels are included in the available pixel
set, leading to the same issue as multiple unmasked cloudy
pixels as described above. The snow/ice and water masks
performed reasonably well. Some instances were discovered
where very turbid water was being included in the mask.
However, within the MIICA process, if any data are masked on
either one of the two dates of imagery, those pixels are exclud-
ed from change detection analysis. Therefore, as long as the
water is masked in one date or the other it does not affect the
change detection results, which generally was the case.
The OLI masks that were developed and enhanced were
generally sufficient, though less so than the LEDAPS masks.
Given the fundamental differences in OLI data compared to TM
and ETM+, including increased radiometric resolution, slight-
ly modified band wavelengths, push-broom system design,
and because the OLI data were corrected to TOA reflectance
rather than SR, the thresholds and spectral tests used for cloud
and cloud shadow masking were modified through several
iterations of processing to tune, as much as possible given
the available time, the algorithm to the data. Occasionally, a
cloudy scene was included in some of the tiled composites.
In these cases, the offending scenes were identified by manual
review, removed from the processing list, and the tile was
then re-processed. In addition, similar issues as above with
cloud edges, thin clouds, and cloud shadows being selected
in the final composites were present in the OLI tiles. For pro-
cessing efficiencies, since the OLI data used a separate prepro-
cessing system, the extent masks were not applied to the OLI
tiles. This did not affect the change detection results since, as
with unmasked water, the previous year’s data were masked
to the processing extent and therefore the change detection
products kept those areas masked.
Composite Results
In general, the composited tiles provided a clean and useful
image source for LANDFIRE RSLC analysis; an example tile set is
shown in Plate 2. Compared to previous updating using single
scene images that were masked for clouds, cloud shadows,
scan gaps, etc., and not filled, the composited tiles contained
much more data, thereby maximizing the ability to capture
landscape change. The LANDFIRE 2010 update used MIICA and
two individual scenes per year to map annual change from
2008 through 2010 (Nelson et al. , 2013b). Since image pairs
for the years 2010 to 2011 were used to run MIICA for both
LF2010 and LF2012, the number of masked pixels was com-
pared to show the effects of compositing versus using single
scenes. The single scenes from LF2010 were mosaiced using
a maximum value filter in overlapping areas so that areas of
change would take precedence over masked-out areas and
then clipped to the LF2012 tiles to facilitate direct compari-
sons. For the leaf-on image dates (day 175 in LF2012) 71 mil-
lion hectares were masked out due to cloud, shadow, snow/
ice, and water in the LF2010 tiles or an average 9.4 percent
of each tile, compared to 47.3 million hectares in the LF2012
composites or an average of 6.5 percent of each tile. For the
leaf-off image dates (day 250 in the LF2012), 71.4 million hect-
ares were masked out of the LF2010 tiles or an average of 9.6
percent of each tile, compared to 39.6 million hectares in the
LF2012 composites, or an average of 5.7 percent of each tile.
The amount of water and snow/ice pixels should be compa-
rable; thus the differences are largely due to cloud and cloud
shadows that are filled in the composite data and not in the
single scenes.
Some data gaps remained in areas with constant cloud cov-
er or perennial snow/ice where no suitable pixels were avail-
able within the requisite time period. In addition, phenologi-
cal changes were evident in some tiles, especially those in the
northern United States with substantial deciduous forests and
cloud cover. In these cases, there were often few valid pixels
available and the acquisition dates varied greatly compared to
pixels in neighboring scenes. This pattern was also noticeable
in agricultural areas primarily in the Midwestern states. An
example generally covering portions of southwestern Michi-
gan, northeastern Illinois, and northern Indiana is shown in
Figure 4 where there is a noticeable scene boundary with phe-
nological variation between scenes. The presence of an area
with no data in the bottom left portion of the image indicates
that scene availability was very low in this area, likely due to
persistent cloud cover near Lake Michigan. Variations in phe-
nology led to mixed results from the MIICA algorithm. Howev-
er, since LANDFIRE does not capture agricultural changes, these
areas were masked in the RSLC process and therefore were not
problematic for LANDFIRE purposes. Reduced resolution mosa-
ics of each tiled dataset were produced, an example of which
is shown in Plate 3.
Change Detection Results
LANDFIRE mapped a total of nearly 17 million hectares of
disturbance for 2011 and 2012 across CONUS. Table 3 shows
the total amount of disturbance mapped by LANDFIRE in CONUS
T able 2. Average Number of Scenes Used to Create Each Composite T ile
Year Target Date Average Number of Scenes
2010 175 54
2010 250 59
2011 175 56
2011 250 62
2012 175 60
2012 250 60
2013 175 61
2013 250 68
580 July 2015 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING