The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
forest,building,mountanious region and so on.In the recent
years,land cover changes are more frequent because of activities
of human beings in Jiaxing. Four cloud-free panchromatic and
multispectral SPOT5 and IKONOS images were acquired over
four years and are considered as our multidate data set. The
acquisition date of these images were July 17th 2002(SPOT-5
pan),August 24th 2002(SPOT-5 multispectral), August 10th
2006(IKONOS pan) and July 18th 2006(IKNONS
multispectral).The spatial resolutions of above images are
different,2.5m for SPOT-5 pan,5m for SPOT-5 multispectral,lm
for IKONOS pan and 4m for IKONOS
multispectral,respectively.
Two preprocessing steps were required for a meaningful
comparison of the satellite images. First, coregistration and
fusion operations between the panchromatic and multispectral
images of the same date were carried out. And then images from
different sensors were registrated to each other with high
precision to avoid misregistration errors inducing false change
alerts.Depending on the satellite image, a set of 49GCPs spread
over the whole study area were selected from satellite images. 3
Figure 1. Fused images left:SPOT-5(2.5m) right:IKONOS(lm)
3. OBJECT-LEVEL METHODOLOGY
OLCD is based on object-oriented analysis technique,the
principle of which is post classification comparison
method.Firstly, Image segmentation partitions an image into
groups of pixels, hereafter named as objects that are spectrally
similar and spatially adjacent, by minimizing the within-object
variability compared to the between-object variability.Secondly,
we classify those objects to thematic map and outline the
polygon objects(vector data format).And lastly,we apply overlay
analysis to objects and estimate whether change or not and
further change type according to the object class attributes.
Figure 2. The flowchart of OLCD based on MS&RQSVM and
OOA
3.1 MEAN SHIFT& REGION GROWMULTISCALE
SEGMENTATION
Objects of interest typically appear on different scales in an
image simultaneously. The extraction of meaningful image
objects needs to take into account the scale of the problem to be
solved. Therefore the scale of resulting image objects should be
free adaptable to the scale of task.Multiscale segmentation aims
to analysize objects in different scale levels.Generally
speaking,in this algorithm,we firstly operate MS procedure on
the images which is the density estimation-based
non-parametric clustering approach.In subsequent steps, region
grow technique merged smaller image objects which acquired
from MS procedure into bigger ones. Throughout this iteractive
clustering process, the underlying optimization procedure
minimizes the weighted heterogeneities (hs, hr and M ) of
resulting image objects.
The multiscale segmentation algorithm(Fig.3) as follows:
Given Xj, Zj(/=1,2, ,«)are J-dimentional space input and
filtered image pixels in spectral and spatial domain,!, is the ith
labeled pixel after segmentation,and hs, hr and M are bandwidth
in spatial and spectral domain and minimum merging region.
(1) Input image data set and transform color space from RGB to
LUV,
(2) Carry out MS procedure in LUV color space and store all
convergent data points to Z\,
(3) Describe the clustering {Cp} p=l,2, m,connect all of the
zj which are smaller than hs and hr in spectral and spatial
domain,
(4) LrlpZjzCp),for i=l,2, n,
(5) Merge adjacent objects according to the parameter M and
obtain segmented images.
Figure 3. The flowchart of multiscale segmentation algorithm
3.2 SUPPORT VECTOR MACHINE CLASSIFICATION
SVM is a useful technique for data classification,which aims at
seperating classes with an optimal hyperplane to maximize the
margin among them (see Fig. 4).It is based on
Vapnik-Chervonenkis (VC) dimension theory and Structural
Risk Minimization (SRM) rule, which solves sparse sampling,
non-linear, high-dimensional data, and global optimum
problems.Its performance has been proved as good as or
significantly better than that of other competing methods in