959
2008
OBJECT-LEVEL CHANGE DETECTION FOR MULTI-TEMEPORAL
HIGH-RESOLUTION REMOTE SENSING IMAGERY
Mu H. Wang a b , Ji.X Zhang 3 , Hai T. Li a , Hua L.Xu ab
a China Academy of Surveying and Mapping,Beijing,l 00039,China-(zhangjx, lhtao)@casm.ac.cn
b Liaoning Technical University ,School of Geomaticas,Fuxin, 123000,China- amudc@126.com,
hualong_0795@ 163 .com
Commission VII, WG VII/5
KEY WORDS: Change Detection, HR images, Pixel-level, Object-level, Segmentation, Classification
ABSTRACT:
A new object-level change detection(OLCD) approach, combining object analysis with change detection process is proposed for land
surface monitoring. The object analysis is consisting of Mean Shift(MS)&Region Grow(RG) multiscale segmentation, Support
Vector Machine(SVM) classification and object footprint. Change detection process is composed of object overlay
analysis(OOA),class attributes comparison and accuracy assessement.Depending on this approach,we can detect the change type of
objects according to classification label. Furthermore,object boundaries are extracted assisted on the vector tool, and the detection
result in the form of vector data can be used to update GIS database in the land use/cover(LUCC) change. The OLCD approach
performances were assessed using multisource SPOT-5 and IKNOS reference data in Jiaxing, and were compared to a pixel-based
method using post classification comparison in CASMImgeInfo3.5.High overall accuracy(>85%) was achieved by object-level
method.The experiment result illustrated the approach could make full use of contextual information of objects and effectively detect
object changes.
1. INTRODUCTION
Remote sensing technique has been widely used in the field of
change detection because of the advantage of macroscopy,high
speed and short interval of acquiring resources,ample
information and effective usability(Massonnet et al., 1993).In
the history of remote sensing applications, many change
detection techniques have been developed.They can broadly
grouped into four categories: visual interpretation
approaches,pixel-level change detection(PLCD),feature-level
change detection(FLCD)and object-level change
detection(OLCD).
Specifically,visual interpretation requires human
experience(computer-assisted or not) to label zones that are
considered as changed,which can make full use of
analysts’experience and knowledge but is time-consuming(Mu
H. Wang et al.,2007).
PLCD approaches extract spectral information of pixels to
describe the geography pattern,and the spatial or contextural
information between proximate pixels is most often
ignored(Atkinson et al.,2000, Townshend et al.,2000).With the
improvement of imagery resolution,the single pixel can’t
represent a region or object but a part.Espectially for urban
areas, the phenomenas that are different object with the same
spetra characteristics and different spetra characteristics with
the same object are severe because of the
materials,furthermore,the effect of projection and shade must be
taken into accounted.All of those make it noninteresting to
analysize object change in the manner of pixel-level
methods(Shackelford A.K., Davis,C.H> 2003; Wang Jianmei et
al.. 2005). Image analysis aims to interpret, quantize and
describe landscape,while the basic unit of landscape is spatially
homogeneous parcel,which composes multiscale interestint
objects.The differences between parcels are
spectra,texture,shape and spatial layout information,which can’t
be provided by a single pixel.
FLCD methods extracted many kinds of features from images
by means of some information extraction techniques such as
Principal component analysis(PCA),texture analysis,shape
analysis,vegetation index,wavelet analysis and so on.And then
we compared those features to decide whether change or
not.Though FLCD has advantages in feature attributes
comparison,while it can introduce other errors in the process of
information extraction.
Recently, pursuers pay more attention on object level image
analysis technique which is similar to visual interpretation.
Instead of analysing pixels independently of their location,
similar contiguous pixels are grouped into objects.The interest
for OLCD methods has increased with the improvements in
image segmentation techniques. The main advantage of
object-based methods is the incorporation of contextual
information in the change analysis (Flanders et al.,2003).
Moreover, the segmentation reduces the local spectral variation
inducing better discrimination between land cover types (Lobo,
1997). However, although the object delineation remains crucial,
a limitation is the definition of a Minimum Mapping Unit
(MMU). Therefore,the final result is largely determined by
objects delineation(Baudouin Desclee et al.,2006).
This research aims to develop a new OLCD methods to detect
land cover and land use change in Jiaxing,Zhejiang Province,
taking advantage of Mean Shift & Region Grow(MS&RG)
multiscale segmentation,Support Vector
Machine(SVM)classification,Object Overlay Analysis(OOA).
This study also aims to test this new approach on a
multitemporal SPOT-5 and IKONOS data set and to compare its
performances to the pixel-level method using the post
classification comparison technique.
2. STUDY SITE AND DATA
The city of Jiaxing covers 5282km 2 and is located in the
southern of Changjiang Delta of China.There are various of land
use type,including agriculture,water body,
Mu H.Wang :Phone : 86 10 88217730; Fax: 86 10 68211420; http:// www.casm.ac.cn