Impervious Comparison of NLCD versus
a Detailed Dataset Over Time
Brandon E. Cooper, Randel L. Dymond, and Yang Shao
Abstract
To address accuracy concerns of the National Land Cover
Dataset (
NLCD
), this case study compares impervious sur-
face from the
NLCD
to a Detailed Land Cover Dataset (
DLCD
)
for the Town of Blacksburg, Virginia over two time periods
(2005/2006 and 2011) at spatial aggregation scales (30 m and
90 m) and scopes (site-specific to area-extent). When compar-
ing the total impervious surface area, the
NLCD
overestimated
significant amounts (12 to 27 percent) for the entire town and
across all specified land use zones (single family, multi-fami-
ly, and non-residential) for both time periods examined. A bi-
nary pixel-wise accuracy assessment of impervious surface re-
vealed that the
NLCD
performed well for the multi-family and
non-residential land use zones. However, accuracy level was
quite low (user’s accuracy <40 percent) for the single family
land use zone. Percent impervious surface of
NLCD
and
DLCD
was further compared at 30 m and 90 m spatial scales. The
spatial aggregation of pixels to 90 m led to improved agree-
ment between the two datasets, although
NLCD
still showed
an underestimate of high values and an overestimate of low
values. An empirical normalization equation was successfully
applied to the
NLCD
to further reduce such data skewness.
Introduction
Imperviousness is an important indicator of urbanization
trends, and can be evaluated through the use of land cover
data which is typically generated from remotely sensed imag-
ery at a variety of spatial scales and scope. Land cover data are
used for a wide range of applications, such as monitoring hy-
drology (Shuster
et al
., 2005), ecological trends (Arnold
et al
.,
1996), population distribution (Wu and Murray, 2007), pol-
lutant contamination (Shields
et al
., 2008), and local climate
change (Yuan and Bauer, 2007). In the US, the Multi-Resolu-
tion Land Characteristics (
MRLC
) Consortium
.
gov/
) generates a National Land Cover Dataset (
NLCD
) at 30 m
spatial resolution using Landsat satellite imagery. Currently,
national-scale land cover data exists for 1992, 2001, 2006, and
most recently 2011. This dataset includes three layers: percent
impervious surface, thematic land cover, and percent tree
canopy. For this study, the
NLCD
percent impervious surface
product is utilized. In this product, estimates of percent (0
to 100) impervious surface are provided for each 30 m pixel.
Detailed land cover datasets are also developed by some local
governments, universities, and research institutions from the
digitization of high spatial resolution aerial imagery; these
data are beneficial for local applications such as land use
planning, watershed management, stormwater infrastructure
design, and water quality regulations (Civco and Hurd, 1997).
The accuracies of the
NLCD
impervious surface maps are
location dependent and frequently do not correctly portray
the heterogeneity of local geographic surface conditions, es-
pecially in urban environments (Homer
et al
., 2004). Studies
suggest that the 30 m spatial resolution might be too coarse to
accurately depict the highly fragmented and spatially com-
plex nature of residential areas in urban and suburban land-
scapes (Giner and Rogan, 2012). For small towns, impervious
surfaces often cover small tracts of land, and their total spatial
extent is relatively small; the data registration accuracy
requirement is significantly high. Increases in data mis-regis-
tration will cause larger uncertainties in land change analysis
or cross-sensor data comparison (Townshend
et al
., 1992).
Other important factors that influence
NLCD
accuracy are: the
capability of available spectral bands to classify specific land
cover types, the radiometric sensitivity of the data, the spatial
resolving powers of the sensors, and the subsequent process-
ing of the data (Mannel
et al
., 2006; Rogan
et al
., 2008).
The
MRLC
reported accuracy statistics mainly rely on
cross-validated estimates of the
NLCD
product versus high
spatial resolution imagery (e.g., 1 to 2 m resolution) from
sources such as the National Agricultural Imagery Program
(
NAIP
). Specifically, the training data for image classification
are randomly divided into two sets. One set is used for the
training of image classification (approximately 1,500 pixels
per
NLCD
mapping zone), and the remaining set is held out
to evaluate the classification performance. Nationwide, the
overall accuracies of 2001 and 2006
NLCD
impervious surface
maps were 79 percent and 78 percent, respectively (Wickham
et al
., 2013). As a result of the national-scale mapping effort,
NLCD
map products are useful for regional or national applica-
tions and are not recommended for projects less than 10 mi
2
(Vogelmann
et al
., 2001).
Because
NLCD
data are readily available for the entire US
for multiple time periods, researchers and local government
agencies are still interested in whether such national-scale
data can be used to support local applications (e.g., stormwa-
ter management). To validate accuracy concerns of the
NLCD
,
locally developed, high spatial resolution impervious surface
maps are needed. Due to cost issues, high resolution data
are not typically developed and such accuracy assessments
have not been routinely conducted (Wickham
et al
., 2013).
Recently, Giner and Rogan (2012) conducted a thorough
comparison of land cover maps derived from Landsat (30 m
spatial resolution) and high resolution imagery (<1 m spatial
resolution). Land cover mapping using high resolution imag-
ery is still a challenging and time-consuming task. Addition-
ally, there may be a temporal discrepancy between
NLCD
and
Brandon E. Cooper is with Greeley and Hansen, LLC,
and previously with Virginia Tech Department of Civil &
Environmental Engineering, P.O. Box 732, Bowling Green,
Virginia 22427 (
).
Randel L. Dymond is with Virginia Tech Department of Civil
& Environmental Engineering, 200 Patton Hall, Blacksburg,
Virginia 24061.
Yang Shao is with Virginia Tech Department of Geography,
115 Major Williams Hall, Blacksburg, Virginia 24061.
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 6, June 2017, pp. 429–437.
0099-1112/17/429–437
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.6.429
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2017
429
