The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
1160
or ecologically homogeneous land class. One reason that the
object-based methods perform well in classifying high
resolution images are because once the object is created by a
segmentation approach, many more features such geometrical
(e.g. shape and area) and topological properties (e.g.
relationship between objects) can be extracted from the
segmented image. This feature is particularly useful in
classifying high spatial resolution images since high spatial
resolution images often contain relatively fewer spectral bands
(e.g. IKNONOS, QIUCKBIRD) compared to coarser images
(e.g. MOIDS, Landsat TM). Consequently, methods that rely on
only spectral information could have difficulty in distinguishing
spectrally similar classes such as buildings and roads). Yet, it
will be much easier to differentiate a building from a road if we
can incorporate the object shape into the classification process.
Hence, in this study, we propose to develop an integrated
approach based on one-class SVMs and object-based methods
to classify one land class from high spatial resolution images.
We first segmented an image by a segmentation approach; both
spectral and spatial properties were then extracted from objects,
the one-class SVMs were then applied to extract one land type
based on properties extracted from the objects. We also
performed the comparisons among the proposed method and the
one-class SVM with pixel-based classification. The overall
accuracy and Kappa coefficient were calculated and used in the
comparison (Congalton and Mead, 1983).
2.2.2 Features extraction: After image segmentation, we then
extract features to be used for image classification. One
advantage of the object-based method is the ability to extract a
wealth of features that could aid in classifying the imagery.
Fourteen features are chosen in this study, detailed descriptions
are described as follows:
(1) Mean value (MV), which represents mean brightness value
of every image object. Since the aerial photo includes three
bands (i.e. green, blue, and red), we will have three mean
values as features for image classification. The formula is as
follows:
¿*=2X ln
;=i
Where, L k is the mean brightness value; n is the number of
pixels in the image object; B ik is brightness value of /th pixel
contained in the image object in band k.
(2) Mean difference to scene (MDS), which is the difference
between mean brightness value of an image object and mean
brightness value of the whole scene in band k. The formula is as
follows:
Sk = IX /n ' IX lm
;=1 j=\
Where, S k is mean difference to the band k; m is the number of
pixels of the whole scene. Similarly, there are three features
which exist in the mean difference to the scene.
2. DATA AND METHODS
2.1 Data
The high resolution remote sensing data used in this research
are from aerial photos with 0.3 meter spatial resolution. The
study area is located in Oakland, California (Figure 1).
Figure 1. Aerial photograph of the study area (color image with
0.3 meter spatial resolution)
2.2 Methods
(3)Mean difference to neighbour (MDN), which is the
difference between mean brightness value of an image object
and mean brightness value of its direct neighbours in band k.
The formula is defined as follows:
1=1 /=1 j—l /=1
Where, p is the number of direct neighbours; mi is the number
of pixels of the neighbour /. There are three MDNs.
2.2.1 Segmentation method: Segmentation methods are used
to generate image objects for classification and image retrieval,
the object is defined as a group of spectrally similar contiguous
pixels. Numerous algorithms have been proposed to segment an
image. In this study, we used the segmentation method from
Defmies software, which is based on a multi-resolution
segmentation algorithm. The segmentation results are tuned
based on scale parameters, color, smoothness, and compactness.
The final segmentation results are shown in Figure 2.
(4) Standard Deviation, which represents the standard deviation
of brightness value of all the pixels contained in an image
object in bank k. There are also three standard deviations.
(5) Area, which represents the number of all the pixels contained
in an image object.
(6) Shape Index (SI) describes the smoothness of the image
object borders, which is useful in differentiating houses and
other man-made objects such as road. The definition of SI is: