PE&RS July 2015 - page 574

Background
Using Landsat-7 ETM+
Several previous studies have investigated the utility of ETM+
SLC-Off data for change detection and other applications.
Variations of two basic approaches are presented in most SLC-
Off studies: interpolation of missing data values, and multi-
temporal scene compositing, to fill in the SLC-Off gaps (Wulder
et al., 2008; Alexandridis et al., 2013; Storey et al., 2005). For
LF2012, there was a preference for using actual pixel values
for change detection, and therefore the interpolation methods
were not considered. Previous compositing approaches used
similar methods for preparing the imagery while varying the
compositing function, or selection of which pixel out of the
available collection, was to be used in the final image. ETM+
compositing methods can be quite useful at filling SLC-Off
gaps given the side to side overlap of adjacent paths, the exact
amount of which varies by latitude. Since the SLC-Off gaps
occur on the edges of each image, using adjacent paths to com-
posite in these areas increases the amount of data available for
gap-filling (Wijedasa et al., 2012). The data available in over-
lapping paths are also often closer in date to the target image
than the next available same-path acquisition 16 days later.
Compositing Methods
One of the most common compositing methods in use has
been the maximum Normalized Difference Vegetation Index
(NDVI) method, which selects the pixel with the highest NDVI
value from the available images, which reduces cloud con-
tamination and other artifacts over vegetated areas (Flood,
2013). The maximum NDVI compositing method has been
used successfully with ETM+ data for deriving multi-temporal
datasets over CONUS and Alaska (Roy et al., 2010). Maximum
NDVI composites are, by design, biased in favor of the “green-
est” vegetation pixels, which may not be optimal for all
applications (Flood, 2013), including change detection where
vegetation disturbance could be masked by a maximum NDVI
composite. For example, if a forest harvest event that removes
green vegetation occurs during a composite period, the pre-
harvest NDVI will be higher than the post-harvest, and there-
fore the disturbed pixel values will be discarded in favor of
the pre-disturbance pixels. Several other pixel-based compos-
iting methods have been proposed that may be more suitable
for vegetation discrimination and change detection mapping.
Flood (2013) used the medoid, or multi-dimensional median
approach to develop seasonal composites that were more
representative of the variability inherent in the time series.
Griffiths et al. (2013) developed an algorithm based on annual
suitability, seasonal suitability, and distance to clouds for large
area land cover mapping. Potapov et al. (2011) used single
band statistics to produce multiple composites from the same
dataset including band-4 median value, and band-5 minimum,
maximum, and quartile (25 percent and 75 percent) values
for mapping forest cover and change. Potapov et al. (2012)
produced start and end date image composites for six-year
time intervals, along with multi-temporal metrics, for clas-
sifying forest cover and change. The image composites were
developed using the per-band median reflectance of the three
earliest and latest observations that passed a series of quality
assessment (QA) analyses and date selection criteria. Hansen
et al., (2013) used these methods as a prototype to map forest
cover and change globally using ETM+ imagery. Each of these
approaches proved more suitable for their intended applica-
tions than the traditional maximum NDVI composite.
Data Masking
Numerous approaches for automated masking of clouds and
cloud shadows in Landsat data have emerged in recent years
(e.g., Irish et al., 2006; Huang et al. , 2010b; Roy et al. , 2010;
Scaramuzza et al., 2012). For vegetation change detection, ad-
ditional features are often discarded, including snow, ice, and
water. Few automated masking algorithms exist that handle
all undesired features for change detection. Two that do ex-
ist are Function of Mask (FMask; Zhu and Woodcock, 2012)
and the Landsat Ecosystem Disturbance Adaptive Processing
System (LEDAPS ; Schmidt et al., 2013). Both algorithms were
designed to work with TM and ETM+ data, though FMask has
since been updated to work with OLI data. OLI standard prod-
ucts do come with a QA band that includes confidence scores
for several types of features, such as clouds (Scaramuzza et
al., 2012), however, not all of the algorithms are currently
implemented, such as cloud shadow detection.
Change Detection
LANDFIRE RSLC processes require an automated algorithm
for detecting annual landscape change using Landsat imag-
ery. While several such approaches exist (e.g., Huang et al.,
2010a; Kennedy et al., 2010; Hansen et al., 2014), LANDFIRE
uses a variation of the Multi-Index Integrated Change Analy-
sis (MIICA) algorithm (Fry et al., 2011; Jin et al., 2013). MIICA
uses pairs of imagery with similar phenology from two time
periods and calculates differences of several indices. Thresh-
olds are used to determine significant changes among the
indices which are then summarized as areas of increasing or
decreasing biomass, or no change. MIICA is run independently
for each pixel in an image and therefore the algorithm is not
affected by using a single scene or composite of multiple
scenes, though the results could be affected by differences in
image acquisition times.
Atmospheric Correction
In previous LANDFIRE updates, data were converted to surface
reflectance (SR) using the LEDAPS algorithm (Masek et al.,
2006; Schmidt et al. , 2013; Nelson et al., 2013b; Maiersperger
et al., 2013). The LEDAPS atmospheric correction algorithm has
been used previously to correct Landsat images from differ-
ent sensors and time periods for building time series stacks
(Huang et al. , 2009) and composites (Griffiths et al., 2013).
However, the SR corrections for OLI data are not yet opera-
tional. Therefore, for LF2012, LEDAPS was used to correct TM
and ETM+ data to SR , and OLI data were corrected to top of
atmosphere (TOA) reflectance using the algorithm published
by the Landsat team(http://landsat .USGS.gov/Landsat8_Us-
ing_Product.php).
Methods
Development of Tiling System
The LF2012 tiling system consists of a grid of 10,000 × 10,000
30 m pixel tiles and is based on the Albers Equal Area Conic
projection for CONUS (Deetz and Adams, 1945; Snyder, 1987)
using standard parallels of 29.5 and 45.5 degrees North, cen-
tered on 23 degrees North and 96 degrees West, and based on
the North American Datum of 1983. This is the same projec-
tion definition used by the National Land Cover Database
(NLCD) project (Chander et al., 2009). Pixels are centered on
multiples of 30 m (Landsat pixel size) within the projection
system. The tile size was selected because a single-scene
Landsat image when re-projected to the Albers Equal Area
projection has a considerable amount of fill and results in a
working raster size of just fewer than 10,000 by 10,000 pixels
(Figure 1). Because analyst workstations and storage systems
were optimized for this size of data from previous updates,
the move to 10,000 by 10,000 pixel, non-overlapping, tiles
was logical. The resulting LANDFIRE tile system is shown in
Figure 2. While most tiles remain 10,000 by 10,000 pixels,
some were extended to capture adjacent coastlines and oth-
ers were cropped to reduce image fill outside the processing
574 July 2015 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING