High-resolution Land Cover and Impervious
Surface Classifications in the Twin Cities
Metropolitan Area with NAIP Imagery
Philipp Nagel and Fei Yuan
Abstract
High-resolution land-use and land-cover (
LULC
) and impervious
surface maps were traditionally created by time-consuming field
surveys or aerial photo interpretation and digitizing methods. In
the United States, the recent availability of 1 m digital color or-
thoimagery through the National Agriculture Imagery Program
(
NAIP
) offers an opportunity to update land cover maps for areas
of interest. In this study, an object-based feature extraction and
a regression tree technique were applied to the
NAIP
imagery
for the Twin Cities Metropolitan Areas (
TCMA
) of Minnesota. An
overall accuracy of 74 percent and 95 percent was achieved for
the general
LULC
classification and the impervious surface map,
respectively. It was found that both techniques could be used to
extract impervious surfaces from the
NAIP
imagery with relative-
ly high accuracy. The
LULC
classification results were compared
to the 2006 National Land Cover Database (
NLCD
) and another
Landsat-based 2006 classification. Implications for future clas-
sification studies were also discussed.
Introduction
Urban land-use and land-cover (
LULC
) and impervious surface
maps have been utilized extensively to assess influences of
urbanization in various previous studies (Schueler, 1994;
Squires, 2002; Dougherty
et al
., 2004; Yuan and Bauer, 2007).
Impervious surfaces, which include building rooftops, streets,
highways, parking lots, sidewalks, and compacted soil or
gravel, accelerate the movement of runoff and pollutants
collected over large area. Therefore, they have been found to
be a “key environmental indicator” to estimate other factors
such as water quality and the overall ecological footprint of
societies (Arnold and Gibbons, 1996; Sutton et al., 2009).
To date, digital remote sensing based
LULC
and impervious
surface classifications over large areas were mainly based on
30 m Landsat data. Medium resolution images offer possi-
bilities for monitoring land-use change and their impacts in
a synoptic view at a regional scale. However, they may not
provide satisfactory results for local level environmental anal-
yses. To extract more detailed
LULC
and impervious surface
maps from the medium resolution imagery, many sub-pixel
classification approaches, including multi-layer percep-
tron model, spectral mixture analysis, regression modeling,
artificial neural networks, and expert systems, have been
developed in previous studies (Yang
et al
., 2003; Dougherty
et
al
., 2004; Wu, 2004; Lee and Lathrop, 2006; Yuan
et al
., 2008;
Yang and Zhou, 2011). In addition to optical remote sensing
images, multi-source ancillary datasets, such as census and
parcel data, surface temperature data, and radar data (Mesev,
1998; Lu and Weng, 2006; Saatchi
et al
., 1997), were also used
in some studies to improve the overall classification accuracy.
High spatial-resolution land cover classifications based on
satellite remote sensing images over large areas are still limited
due to the high cost of the data. The recent availability of 1 m
resolution National Agricultural Imagery Program (
NAIP
) imagery
in the United States from the Geospatial Data Gateway of Natu-
ral Resources Conservation Services (
NRCS
) and advancement of
geospatial technologies offer an opportunity to generate regional
high-resolution
LULC
and impervious surface maps. Therefore,
the overall objective of this study was to provide an example for
extracting high-resolution
LULC
and impervious surfaces from
the freely available
NAIP
imagery in the Twin Cities Metropolitan
Area (
TCMA
) of Minnesota. First, an object-based classifier was
used to derive general
LULC
classes. Next, a decision-tree model
was implemented to extract impervious surface areas based on
a combination of data sources. Further, results were analyzed
and compared to two Landsat-based classifications for this study
site. The techniques, accuracies, and implications for future
classification studies were also discussed.
The increased within-class spectral variability is a major
driver for applying object-based image analysis to high spatial
resolution imagery. Object-based image analysis not only
takes into account spectral information, but also other spatial
and contextual factors such as the object’s shape, size, texture,
pattern, elevation, and spatial relationships with neighboring
objects. Therefore, it has proven to be an effective approach
for high spatial-resolution image classification (Al-Khud-
hairy
et al
., 2005; Antonarakis
et al
., 2008; Bhaskaran, 2010;
Blaschke, 2010; Whiteside
et al
., 2011; Nagel
et al
., 2014).
A decision tree method was used to extract the impervious
surfaces based on the literature review of previous research.
Decision tree modeling is a machine learning technique that
analyzes existing data and fits the data best into predetermined
classes through a tree model (Breiman, 1984; Wu and Kumar,
2009). The method provides several advantages over tradi-
tional classification methods such as maximum likelihood
classification and minimum distance classification. In par-
ticular, it offers more flexibility for the requirements of input
data. It can use continuous and thematic data as inputs and
can handle multi-source data with different scales and units.
It does not depend on statistical assumptions about the data
and can process large data sets quickly (Hansen
et al
., 1996;
Gahegan, 2003; Pal and Mather, 2003). Due to these advantag-
es, the method was adopted to generate the 1 km global land
Philipp Nagel was with the Department of Geography, Minne-
sota State University, Mankato, MN 56001, and currently with
the Sibley County Public Works, 111 8
th
Street, P.O. Box 897,
Gaylord, MN 55334.
Fei Yuan is with the Department of Geography, Minnesota
State University, Mankato, MN 56001
).
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 1, January 2016, pp. 63–71.
0099-1112/16/63–71
© 2015 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.1.63
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
January 2016
63
