PE&RS July 2015 - page 583

Comparisons were also made with WELD composites to
illustrate the differences between compositing methods.
WELD summer composites for 2010 and 2011 were re-tiled
to the LF2012 tiling system and used to run the LF MIICA
change detection algorithm. Plate 4 shows 2011 leaf on tiled
imagery from LF2010, LF2012, and WELD with 2010 to 2011
leaf on MIICA results for tile “r7c11,” which covers portions
of Mississippi, Alabama, and Tennessee. Visible differences
exist between the images and MIICA results. The most obvious
differences are the amount of masked pixels (white) in the
LF2010 MIICA results, scan line artifacts in the WELD imagery
and MIICA output, and the amount of change detected by MIICA
using each of the three image sources from the same time pe-
riod. LF analysts who manually processed every LF2010 scene,
and also processed each LF2012 tile indicated a qualitative
increase in the ability to detect and classify landscape change
using the LF2012 process. Visual inspection of the 2010 to
2011 WELD MIICA results, in comparison with the LF2012 MIICA
results, showed a general increase in the amount of change
mapped in the LF2012 data, with more anomalous change
flagged in the WELD data, as illustrated in Plate 4.
Discussion and Conclusions
Advantages to Tiled Approach
Moving from the previous individual scene-based processing
to tile-based processing resulted in several advantages. Scene
selection was able to be fully automated, compared to previ-
ous updates, which required manual selection of scenes to
achieve the best combination of seasonally matching scenes
of greatest image quality. Even with trained analysts viewing
every available scene, there were many cases where the best
available scenes had significant cloud cover and seasonality
did not match well between image dates because of limited
scene availability.
By using the best portions of every available scene, the
composited tile approach resulted in scenes of greater image
quality with more tightly controlled seasonal dates, reduc-
ing the impact of phenological changes between years. The
number of images to be processed post-MIICA was reduced
by 78 percent; whereas in previous LANDFIRE updates, 445
individual scene frames were mapped by analysts, only 98
tiles were needed to cover the same area. The size of each tile
was only slightly larger than the size of the re-projected single
images, though there were more data within each tile, the fill
area being replaced with pixels from neighboring scenes to fill
the entire tile extent with valid data, each pixel only needed
to be mapped once, eliminating the redundant processing of
overlap areas.
The results shown in Plate 4 were similar for other por-
tions of the country. In general, the LF2012 change detection
results had much less area masked out due to clouds and
cloud shadows than LF2010, allowing more change to be
detected. Conversely, when no reasonable data existed for a
given pixel, a no-data value was assigned, preventing anoma-
lies in the change detection results as seen in the WELD data
where all pixels are given valid data, even if all available
images are cloudy.
Limitations
Several limitations in the developed tiling and composit-
ing approach led to decreased image quality in some areas.
The use of static target dates across the entire country likely
contributed to the phenology-related issues described above.
A more reasonable approach might be to define regional target
dates that consider phenology patterns of local vegetation
and cloud dynamics to optimize the time periods desired for
compositing. For example, no scenes were selected between
days 1 to 99 or days 300 to 365 (or 366). In some areas, espe-
cially the southeastern United States, images from these time
periods may be appropriate. Adjusting the time period for a
particular region would require only simple modifications
to existing processing scripts to implement, once the desired
date ranges were identified.
Improvements to the cloud, cloud shadow, water, and
snow/ice masks could reduce the prevalence of undesired
data in the final composites. The LEDAPS masks were chosen
for the LANDFIRE process based on an evaluation of several
different automated data masking routines that found the
Plate 3. 2012 day 175 composite mosaic.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING July 2015 583