348
May 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
as an upland vegetation class was changed to a wetland class if
all three ancillary datasets identified it as a wetland class). For
major forest transition areas in the southeastern and north-
western U.S. regions, applying knowledge about forest distur-
bance, succession stage, and management practice patterns in
conjunction with a spectral-ranking approach greatly improved
the quality and consistency of land cover change labeling by
preventing illogical land cover changes in the final product.
U
rban
I
mpervious
E
stimation
The approach for updating new impervious surface growth
and intensification between 2006 and 2011 was similar to
the method used to produce the 2006 NLCD impervious sur-
face change product (Xian et al., 2011; Xian et al., 2012). This
method employed the NLCD 2006 impervious surface product
as the baseline estimate and Landsat imagery pairs in 2006
and 2011 as the primary data source for identifying changed
areas. Ancillary data such as DMSP NSLS, slope, and eleva-
tion were also used to help develop regression tree models for
predicting new percent impervious surface in changed areas.
Three major steps were required for this process: (1) modeling
an impervious surface, (2) comparison of model outputs, and
(3) final product clean-up. In step 1, DMSP nighttime lights
imagery in 2006 was superimposed on the NLCD 2006 imper-
vious surface product to exclude low density impervious ar-
eas outside urban and suburban centers to ensure only urban
core areas be used to provide a stable and reliable training
dataset. Two training datasets, one having a relatively large
urban extent and one having a relatively small extent, were
produced through imposing two different thresholds, ≥ 10
and ≥ 20, of nighttime lights imagery on the 2006 impervious
product. In step 2, each of the two training datasets combined
with 2006 Landsat imagery was separately applied with re-
gression tree algorithms to build up regression tree models
(Xian and Homer, 2010). Two sets of regression tree models
were created and used to produce two 2006 synthetic impervi-
ous surface products. Similarly, the same two training data-
sets were used with 2011 Landsat and DMSP
NSLS images to create two sets of regression
tree models and produce two 2011 synthetic
impervious surface products to ensure that
only stable predictions are chosen as inter-
mediate products. In step 3 the two synthetic
product pairs were then compared to remove
false estimates due to strong reflectance from
nonurban areas and to retain 2006 impervious
values in the unchanged areas. The 2011 im-
pervious surface was updated individually in
every Landsat scene over the entire CONUS,
with individual scene products subsequent-
ly mosaicked together to produce a seamless
2011 impervious surface product.
In addition to identifying new impervious
features for 2011, the process was sensitive
enough to capture many previously unidenti-
fied impervious areas from earlier periods. Identifying these
areas as 2011 change would have inaccurately placed the
change in the wrong period. To correct this, an intensive com-
bination of hand editing and automated processes was applied
to identify and sort potential additions into the proper NLCD
period (2001, 2006, or 2011). This approach was dependent on
extensive use of high-resolution imagery from each period to
accurately identify and sort the additions. All other impervi-
ous features were also checked during this process, enabling
overall accuracy to be improved. These special edits were fo-
cused on the eastern half of CONUS because this area had
the most inaccuracies from earlier periods. The additional
processing resulted in a much improved impervious product
throughout all published years and a more consistent nation-
al product.
T
ree
C
anopy
C
over
The NLCD 2011 percent tree canopy cover was modeled using
photographic interpretation of National Agriculture Imagery
Program (NAIP) aerial imagery, Landsat 5 imagery (and deriv-
atives), ancillary data such as elevation (and derivatives), and
previous NLCD data (Fry et al., 2011). Approximately 65,000
sample locations were photo-interpreted for percent tree cano-
py cover using NAIP imagery. These data and corresponding
Landsat and ancillary data were used to develop a random for-
est model (Brieman, 2001) for each NLCD mapping zone (Ho-
mer and Gallant, 2001). Two versions of the NLCD percent
tree canopy cover were developed: an analytical version and
cartographic version. The analytical version is intended to be
used for estimating average tree canopy cover in an area of in-
terest and includes both percent tree canopy predictions and
uncertainty around those predictions (Coulston et al., 2014).
The cartographic version is intended to be used more as a visu-
al backdrop in cartographic applications. Masking procedures
and other post-processing procedures were used to reduce com-
mission error and smooth seamlines between mapping zones.
The 2011 NLCD percent tree canopy cover differs from the
