Comparison of Three Landsat TM Compositing
Methods: A Case Study Using Modeled
Tree Canopy Cover
Bonnie Ruefenacht
Abstract
Landsat imagery mosaics developed using model II regression
have been shown to successfully model percent tree-canopy
cover (
PTCC
). Creating model II regression mosaics, however, is
a time-consuming, manual process. The objective of this study
is to evaluate the effectiveness of using more easily automated
image composites techniques, such as median Landsat-5 image
composites or maximum
NDVI
Landsat-5 image composites, as
alternatives to model II regression mosaics for the modeling
of
PTCC
. This study found all composite types were effective in
modeling
PTCC
, but the maximum
NDVI
composites included
anomalies, clouds, shadows, and tended to be pixelated, where-
as the median composites and the model II regression mosaics
had none of these issues. The median composite procedure is
automated and was found to be an effective approach to statis-
tically reduce a much larger ensemble of images on a pixel basis
to create images suitable for vegetation modeling.
Introduction
A 30-meter raster Landsat-based percent tree-canopy cover
(
PTCC
) dataset for the conterminous United States, coastal
Alaska, Hawaii, and Puerto Rico was first produced as part of
the 2001 National Land Cover Database (
NLCD
) (Homer
et al.
,
2007). A
PTCC
dataset was also produced as part of
NLCD
2011
(Jin
et al.
, 2013). Currently, the 2011
PTCC
dataset only covers
the conterminous United States (
CONUS
), but will eventually
be developed for coastal Alaska, Hawaii, and Puerto Rico.
Before developing the
NLCD
2011
PTCC
dataset, a pilot study
investigated the feasibility of mapping
PTCC
using different
datasets and model algorithms (Coulston
et al.
, 2012). Five
areas, located in Georgia, Kansas, Michigan, Utah, and Ore-
gon, were chosen as pilot study areas; each area being slightly
larger than the size of a single Landsat scene (183 km × 170
km) and centered over the intersection of four Landsat World-
wide Reference System (
WRS-2
) paths/rows allowing for the
investigation of Landsat seam-lines on the modeling of
PTCC
.
The response data used by Coulston
et al
.
(2012) were
USDA
Forest Service Inventory and Analysis (
FIA
) data sampled
at different intensities and photo-interpreted for
PTCC
. The
explanatory data consisted of radiometrically non-normalized
Landsat-5
TM
(L5) data, radiometrically normalized L5 data,
digital elevation model (
DEM
) data, and various derivatives of
these datasets. The non-normalized L5 mosaic was created by
converting each L5 scene to top-of-atmosphere and then to
surface reflectance using dark-object subtraction and mosa-
icked using a simple overlay. The normalized L5 mosaic was
created by calibrating each L5 scene to adjacent L5 scenes us-
ing model II regression techniques (Beaty
et al.
, 2011). Using
these response and explanatory data, Coulston
et al
. (2012)
found that using 20 percent of the available
FIA
data was an
effective sampling intensity for modeling
PTCC
and that the
random forest algorithm (Breiman, 2001) performed better
than beta regression for deriving
PTCC
estimates. Coulston
et
al.
(2012) also found that there were no significant model per-
formance differences between using normalized images and
non-normalized images.
In addition to investigating datasets and model algorithms,
Coulston
et al
. (2012) also performed a scaled-up prototype
study to develop a final methodology for the
CONUS
modeling
of
PTCC
for
NLCD
2011. The prototype study was conducted on
five Multi-Resolution Land Characteristic (
MRLC
) consortium
mapping zones (Homer and Gallant, 2001), which were zones
16, 23, 48, 54, and 59 (Figure 1). The prototype study used
20 percent of the
FIA
PTCC
data available for those zones and
random forest was the model algorithm. Normalized L5 model
II regression mosaics were used for the prototype because they
created more visually appealing maps that did not have seam-
lines.
The model II regression normalization procedure is not a
computationally time-intensive procedure, but is a time-con-
suming, manual procedure that took two to five person days
per Landsat scene. To create model II regression mosaics for
the 456
WRS-2
Landsat paths/rows for
CONUS
would take three
to eight person years. This cost was deemed to be unaccept-
able for the production of
PTCC
for
NLCD
2011 and, thus, an
alternative had to be found. Image compositing is a method
used to reduce a series of images into a single image. Image
composites are routinely created for Moderate Resolution Im-
aging Spectroradiometer (
MODIS
) images (Wolfe
et al.,
1998).
MODIS
provides full global coverage every two days. To reduce
this volume of data, image composites are created by selecting
the best pixel based upon view angles, absence of clouds and
shadows, and aerosol loading.
Landsat provides global coverage on a 16-day cycle.
Because of the temporal nature of Landsat, Landsat imagery
is not typically composited. When using Landsat images for
mapping or modeling of vegetation characteristics, single-day
or multi-date imagery representing different phenological pe-
riods are typically used (Conese and Maselli, 1991; Schriever
and Congalton, 1995; Wolter
et al
., 1995). Landsat imagery
from three dates representing early, peak, and late vegetation
green-up was used to develop
NLCD
2001 (Homer
et al.,
2004).
These multi-date images were used as is and were not com-
posited into a single date. Coulston
et al.
(2013) investigated
whether multi-seasonal imagery is necessary for the modeling
of
PTCC
and found no significant differences between
PTCC
modeled with multi-seasonal imagery and
PTCC
modeled with
USDA Forest Service Remote Sensing Applications Center,
2222 West 2300 South, Salt Lake City, Utah 84119 (bruefen-
).
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 3, March 2016, pp. 199–211.
0099-1112/16/199–211
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.3.199
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2016
199
