Beruflich Dokumente
Kultur Dokumente
186
KEY WORDS: High Spatial Resolution, Image Segmentation, Pixel-based Classification, Object-based Classification, Mapping
Mechanism
ABSTRACT:
Image classification is a challenging problem of high spatial resolution remote sensing image. On the basis of analyzing and
summarizing the research actuality of remote sensing image classification technology, this paper proposed a new object-based image
classification method based on mapping mechanism for high spatial resolution remote sensing image. The classification framework
used a special mapping strategy to fit in the special data format and content of high spatial resolution remote sensing data. First, the
multi spectral image was segmented by multi scale watershed segmentation and at the same time classified by a traditional pixel-
based classification method (maximum likelihood); then the pixel-based multi spectral classification result was mapped to the
segmentation result by area of dominant principle to get the object based multi spectral classification result. In order to make good
use of the information in the pan image, it was also segmented, and the final classification result was gotten by mapping the object-
based multi spectral classification result to pan image segmentation result. Experiment results show that the mapping mechanism
based classification algorithm for high spatial resolution remote sensing data can make use of the information both in pan and
multispectral bands, integrate the pixel-based and object-based classification method, and finally improve the classification accuracy.
186
integrate the pixel-based and object-based classification method, image is very complex. So its easy to make an over-
and finally improve the classification accuracy. segmentation. In order to obtain a concise region representation,
2. METHODOLOGY post-processing are applied to resolve the over-segmentation.
Post-processing (region merge) merges regions according to
This section details watershed transformation, multi-scale certain criteria for a more concise region presentation.
segmentation, region merging and mapping algorithm, Region merge is an important method to further reduce the
maximum likelihood classification (MLC) applied to multi over segmentation. Two neighboring regions can be merged
spectral high spatial resolution remote sensing image and into a single region if they are similar enough. The region
mapping mechanism integrate the pixel-based classification merging algorithm presented here makes use of the spectral and
result and object-based segmentation result. The flow chart of texture properties of regions as its merging criteria. This section
the proposed object-oriented classification is shown in figure 1. describes the criteria considered in this merging process.
A flow chart of region-merging algorithm is presented in
Multi Spectral Image Pan Image figure 2. After the segmentation through multi scale region
transformation is obtained, the associated regional parameters,
weight of spectral heterogeneity and weight of texture
Multi Scale heterogeneity are applied to an iterative region-merging
segmentation algorithm outlined as follows.
Pixel Based
Classification
Figure.1 The flow chart of proposed classification method Finale Segmentation Result
Figure.2 The flow chart of region merging for multi spectral
2.1 Watershed Segmentation
image
Image segmentation is very essential and critical to image
processing and pattern recognition. Watershed transformation is Traditional region-merging algorithms are based on
a powerful mathematical morphology technique for image spectral heterogeneity. Although such information is abundant
segmentation and subsequently a lot of algorithms for its in high resolution remote sensing images, complications due to
implementation have been proposed. noise and influence among objects are high for these cases as
Watershed algorithm is a good choice for the high well. Merging regions by the spectral criterion alone will yield
resolution image with large amount of data because of its fast suboptimal results. For this study, a mixed criterion was
segmentation speed. The idea of watershed is drawn by implemented that combines both spectral and spatial
considering an image as a topographical surface (John Goutsias, heterogeneity criteria as developed.
2000). Suppose we pierce holes at every regional minimum and
dip the image surface into water, then water will flood areas 2.3 Maximum Likelihood Classification (MLC)
adjacent to regional minima. A regional minimum is a
There are many image classification algorithms for remote
connected plateau from which it is impossible to reach a point
sensing images. Each algorithm has its unique advantage and
of lower gray level by an always-descending path. As the image
weakness. In this paper, we focus on the application of the
surface is immerged, some of the flood areas will tend to merge.
mapping mechanism in high spatial remote sensing images
When two or more different flood areas are touched, watershed
classification. Theoretically, all the pixel-based classification
lines (i.e. dams) are constructed between them. When finished,
method can be applied in the proposed algorithm. Therefore, in
the resulting networks of dams define the watershed of the
order to improve the efficiency of the whole approach and make
image.
it more simple, the maximum likelihood classifier is chosen
We apply the labeling watershed segmentation algorithm
with the advantages of simple and rapid.
to segment the pan and multi spectral image. Morphological
Maximum likelihood classification (MLC) is a method for
operators were used to attract the gradient of the image, which
determining a known class distribution as the maximum for a
provide the watershed segmentation algorithm with labeled
given statistic (Scott and Symons, 1971). MLC is, in fact,
areas.
widely used in remote sensing, in which a pixel with the
maximum likelihood is classified into the corresponding class.
2.2 Region Merging
Suppose there are m predefined classes, the class a posteriori
However watershed algorithm is very sensitive to the noise probability is defined as(Liu et al, 2011)
and detail information, and high resolution remote sensing
187
P(k ) P( x | k ) (1) taking resolution ratio 1:4 for example. One pixel in multi
P (k | x) m spectral image corresponds to sixteen pixels in panchromatic
P(i) P( x | i)
i 1
image.
1 1
exp ( x k )T k ( x k )
1 (2)
Lk ( x)
2
n
(2 ) | k |
2
nn
n1 n 2
matrix of class k.
188
SPOT-5 satellite data were applied to make a series of
experiments and comparative analysis between the mapping
mechanism based classification method and pixel-based MLC
method.
The classification experiment one is performed with the
SPOT5 satellite remote sensing data, including the pan band
with the resolution of 2.5 meters and four multi-spectral bands
with the resolution of 10.0 meters. The size of multi spectral
image is 300300, and 12001200 for pan image.
The classification experiment two is performed with the
IKONOS satellite remote sensing data, including the pan band
with the resolution of 1.0 meter and four multi-spectral images Figure 9. The IKONOS multi spectral image pixel-based
with the resolution of 4.0 meters. The size of multi spectral classification result
image is 256256, and 10241024 for pan image.
189
processing. It can be seen from figure 11 that the proposed classification with SPOT HRV data. Remote Sensing of
object-based classification method can obtain better Environment, 40, pp. 137151.
classification result than MLC result. And the plaques noise is John Goutsias, 2000, Mathematical morphology and its
effectively reduced. Although the post-processing can restrain applications to image and signal processing, Kluwer Academic
the plaques noise in pixel-based classification result, the Publishers.
boundaries of regions have been dislocated. Kimfung Liu, Wenzhong Shi, Hua Zhang, 2011, A fuzzy
topology-based maximum likelihood classification, ISPRS
Journal of Photogrammetry and Remote Sensing 66 103-114
Scott, A.J., Symons, M.J., 1971. Clustering methods based on
likelihood ratio criteria. Biometrics 27 (2), 387397.
Zhong Chen, Guoyou Wang & Jianguo Liu, 2012, A modified
object-oriented classification algorithm and its application in
high-resolution remote-sensing imagery, International Journal
of Remote Sensing, 33:10, 3048-3062.
4. CONLUSION
ACKNOWLEDGEMENT
REFERENCE
190