modu-
rkflow.
remote
rt No.
Profes-
T 1535
ssional
5, CAT
^rofes-
S. and
digital
ety for
laska.
image
n from
Algo-
' qual-
ed im-
rithms
nd se-
fitting.
}5-61.
lkit in
al Re-
Cision
E.
ry and
Com-
ul 2004
IMAGE CLASSIFICATION WITH A REGION BASED APPROACH IN
HIGH SPATIAL RESOLUTION IMAGERY
Jong Yeol Lee *, Timothy A. Warner b
? Korea Research Institute for Human Settlements, 1591-6, Kwanyang-Dong, Dongan-Gu, Anyang-Shi, Kyounggi-
Do, 431-712 Republic of Korea - jylee(@kkrihs.re.kr
" West Virginia University, Department of Geology and Geography, PO Box 6300, Morgantown, WV 26506-6300,
USA - Tim. Warner(@Mail.wvu.edu
KEYWORDS: Image segments, classification, object-oriented approach, multivariate analysis, high spatial
resolution
ABSTRACT:
A number of studies have been carried out to find an appropriate spatial resolution to which to aggregate data in order
to reduce the variation within an object, and minimize the classification error. Such approaches are pixel-based, and
do not draw on the spatial variability as a source of information. The variability within an object can provide
additional information that can be used for image classification. Instead of pixels, groups of pixels that form image
segments, which are called “patches” in this study, were used for image classification. New methods that exploit
multivariate statistics to improve the image classification are suggested. In the case of the object-based classification,
patches are not expected to consist of pixels with completely homogeneous spectral radiances, but rather certain
levels of variability are expected. To treat this variation within objects, multivariate normal distributions are assumed
for every group of pixels in each patch, and multivariate variance-covariance matrices are calculated. A test of this
approach was conducted using digital aerial imagery with a nominal one meter pixel size, and four multispectral
bands, acquired over the small city of Morgantown, West Virginia, USA. Four classification methods were
compared: the pixel-based ISODATA and maximum likelihood approaches, and region based maximum likelihood
using patch means and patch probability density functions (pdfs). For region-based approaches, after initial
segmentation, image patches were classified into seven classes: Building, Road, Forest, Lawn, Shadowed
Vegetation, Water, and Shadow. Classification with ISODATA showed the lowest accuracy, a kappa index of 0.610.
The highest accuracy, 0.783, was obtained from classification using the patch pdf. This classification also produced
a visually pleasing product, with well-delineated objects and without the distracting salt-and-pepper effect of isolated
misclassified pixels. The accuracies of classification with patch mean, and pixel based maximum likelihood were
0.735, 0.687 respectively.
single spatial resolution that suppresses all unwanted
spectral variability (Marceau et al, 1994a; 1994b).
Perfect classification could be achieved if each spectral Studies that use image segmentation to identify single
1. INTRODUCTION
class were to have a unique spectral signature.
However, spectral overlap between most real classes
occurs as a result of noise in the system, the natural
variability of objects within a specific class, and the
spatial variability of radiance within each object (Swain
and Davis, 1978; Price, 1994). An added
complication is that the spectral structure of an image is
a function of scale (Cao and Lam, 1997). Higher
spatial resolution may actually lead to greater
variability within classes, as additional detail 1s
resoived.
A number of studies have been carried out to find an
appropriate spatial resolution to which to aggregate
data in order to reduce the variation within an object,
and minimize the classification error (Pax-Lenney and
Woodcock 1997; Teillet et al, 1997; Latty et al,
1985). Such approaches are pixel-based, and do not
draw on the spatial variability as a source of
information. Another problem with methods that
search for an optimal scale is that real objects and
classes are variable in size, and thus there is usually no
objects (Gougeon, 1995a) can overcome this problem
ofa single optimal scale. However, most such studies
use mainly aggregated information such as average DN,
and to a limited extent the variance within the image
segments (Kettig and Landgrebe, 1976; Gougeon,
1995a; Meyer et al., 1996). The variability within an
object can provide additional information that can be
used for image classification. The spectral correlation
between bands, as quantified by the covariance matrix,
is in fact often a key determinant in traditional
maximum likelihood classification for separating
classes that overlap in their univariate distributions.
2. TRADITIONAL IMAGE CLASSIFICATION
Maximum likelihood classification is a standard, pixel-
based supervised approach, which classifies unknown
pixel-based on multivariate probability density
functions (pdf) of the classes of interest. Statistical
properties of training data sets from ground reference
data are typically used to estimate the pdfs of the