cover classification, the 30 m large area impervious surfaces,
US National Land Cover Database (
NLCD
), and the US Crop-
land Data Layer (Hansen
et al
., 2000; Yang
et al
., 2003; Homer
et al
., 2007; Johnson and Mueller, 2010). Moreover, Yuan
et al
.
(2008) compared three major techniques (regression model-
ing, regression tree, and normalized spectral mixture analysis)
for percent impervious surface area (
%ISA
) estimation using
Landsat imagery for the
TCMA
. They found that the regression
tree method generated the most accurate
%ISA
.
Study Area
The seven-county
TCMA
(Anoka, Carver, Dakota, Hennepin,
Ramsey, Scott, and Washington counties) has a total area of
approximately 7,700 km
2
(Figure 1). It is the eighth fastest
growing area in the US with a population of 2.8 million as of
2010, which comprised approximately 54 percent of the total
population of Minnesota (US Census, 2013). The area is a
major transportation hub for rail and water cargo and passen-
ger services. Tertiary sector businesses, high-tech research and
production, and financial services are today’s major economic
bases. Several major corporations and Fortune 500 companies,
such as Target Corporation, Dairy Queen, 3M, and General
Mills are headquartered in the Twin Cities. From 2010 to 2030,
TCMA
’s population is predicted to increase about 18.5 percent
to 3.38 million (Metropolitan Council, 2014). High-resolution
LULC
information will help improve the ability of assessing the
urbanization influences of
TCMA
on the ecosystem.
Data and Methods
Aerial Imagery and Preprocessing
The 2010
NAIP
orthoimagery with 1 m spatial resolution were
obtained from the Minnesota Geospatial Information Office
(
MNGEO
) Data Clearinghouse
/
).
The images have red (R), green (G), blue (B), and near-infrared
(
NIR
) spectral bands and were mainly collected in July and Au-
gust. They were ortho-rectified, mosaicked, and color-balanced
by the data vendor to produce county mosaics in MrSID
™
(.sid)
format, which allows image files to be substantially compressed
with little to no loss of image quality. The
NAIP
imagery was
first downloaded as full county mosaics for each of the seven
counties in
TCMA
. Next, they were uncompressed to the image
format of ERDAS Imagine
®
(.img) that has full spatial resolu-
tion. The total size of the seven uncompressed county mosaics
was 116 gigabytes. Then, they were further mosaicked using the
Mosaic Dataset tool of ArcGIS
®
, which is a dynamic processing
tool that allows images to be mosaicked on the fly. Finally, the
dynamically mosaicked image was clipped to the
TCMA
. In ad-
dition, to provide additional classification inputs, texture values
were calculated from the
NIR
band as variance of a 3 × 3 pixel
window. Texture information has been found to be valuable for
object-based classifications (Ryherd and Woodcock, 1996; Yuan,
2008). It can also be used in a pixel-based decision tree approach.
Lidar Elevation Data and Road Base Map
The Light Detection and Ranging (lidar) digital elevation
model (
DEM
) data, acquired in 2011 and 2012, were also ob-
tained from
MNGEO
. The 1 m lidar-based
DEM
images were also
mosaicked dynamically and clipped to the
TCMA
. In addition,
a base map of roads was downloaded from the Minnesota
Department of Natural Resources (
MNDNR
) Data Deli (
http://
deli.dnr.state.mn.us/
). Roads from the map are represented
as centerlines. Because road density is often considered an
important measure of urbanization (Schueler, 1994), a road
density raster was computed as kilometers of road for one
square kilometer around each pixel based on the road center-
line layer. The
DEM
and the road density map were later used
as the inputs for the classifications.
General Land-Use and Land-Cover Classification
Classifications were done subsequently on the
TCMA
mosaic
image of
NAIP
. The general
LULC
classification was performed
using Feature Analyst, which is based on a proprietary,
object-based, inductive learning classification algorithm. The
following six classes were extracted: (1) water (rivers, lakes,
pools, and other open bodies of water); (2) urban infrastructure
(roads, buildings, etc.); (3) cropland (non-vegetated and vege-
tated agriculture fields and pasture); (4) forested areas (decidu-
ous, evergreen, and mixed forests); (5) other vegetated areas
(shrub, herbaceous plants, non-forested wetlands, lawns); and
(6) bare soil and rock (mining operations such as gravel pits,
bedrock). Particularly for the cropland class, it was first clas-
sified into vegetated and non-vegetated cropland classes and
then merged to one class in the next step because initial testing
showed that this increases cropland classification accuracy.
Moreover, due to limited spectral resolution of
NAIP
imagery
and the lacking of field data, wetlands, shrubs, and herbaceous
vegetation were classified as one class: “other vegetated areas.”
Training samples that represented both the spectral infor-
mation and the size and shape of the objects, as well as their pat-
terns, texture, and neighboring objects, were determined manu-
ally for each class. They were drawn as close to the edge of each
feature as possible to allow the sample to represent the shape
and edge type of the object. The variance of each class was taken
into account also. The classifier divides the individual pixels
into objects using the training samples in the image segmenta-
tion process. After creating the training polygon set, other set-
tings were defined. For example, multi-source input data were
defined, which included the
NAIP
imagery, the lidar
DEM
, and the
texture layer. To make use of contextual information efficiently,
the foveal input representation for each class was defined,
which allows the classifier to give more importance to pixels
nearer to the pixel being processed (Opitz and Bain, 1999).
Impervious Surface Extraction
Creation of the Decision Tree
To extract impervious surfaces, a decision tree classifier,
See5 software, based on the C5 algorithm was used. The
only difference between the impervious surface map and the
urban class in the
LULC
map is the compacted bare soil on the
Figure 1. The seven-county Twin Cities Metropolitan Area of Minnesota.
64
January 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
