bul 2004 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
anbul 2004 
  
  
step1: 
extraction of 
stationarity 
| Oneisa [m 
atures. The 
ry changes 
ıltural area 
ltural area 
io the stage 
ing. These 
€, and they 
ita. On the 
omly and 
ge can be 
disasters as 
quire data 
he rule for 
ic of each 
tionarity of - stationary change 
of changes > - nonstationary change 
anges. The pod 
for making v 
be updated. 
btraction of 
ever, these 
the changes 
nge can be 
for a long 
yugh. This 
  
  
  
  
step2: 
E, evaluation of 
AR image stationarity 
  
  
        
  
  
  
  
   
  
  
  
  
requirement can not be satisfied by optical sensors which can 
not regularly provide useable images of the land surface. In this 
sense, SAR imagery, which has an assured periodicity, is the 
I most suitable data to satisfy the requirement. 
  
The method developed in this study enabled the discrimination 
of “change” to be detected in the N™ (N > 1) observation, 
and separated into stationary change or nonstationary change 
by comparing the change with the former archived data (1 : N- 
1). The method consisted of two steps. 
In the first step, stationarity of change was extracted by time 
series analysis performed on the multitemporal SAR images 
and a change process model was built into each pixel. A 
disadvantage of using SAR data is its speckle noise, which 
j makes image interpretation difficult. Most of the speckle noise 
1998 reduction filters have a common defect of degrading spatial 
resolution. Principal Component Analysis (PCA) was adapted 
to multitemporal SAR images to reduce the speckle noise by 
maximizing the advantage of temporal information as a test. 
PCA resulted in much clear image compared to the single 
image (refer to Figure 3). In addition, there are several 
multitemporal speckle filters (e.g. Bruniquel and Lopés, 1997; 
Coltuc et al., 2000; Ciuc et al., 2001). But still PCA and most 
of their filters do not use the information of periodical change 
of feature. On the other hand, time series analysis has focused 
on the periodicity, and enabled the extraction of both authentic 
change (i.e. base fluctuation) and the noise (i.e. deviated 
fluctuation). A key distinguishing feature for developing this 
method is in its extraction of trend variations; cyclic and 
| seasonal variations; and irregular variations. Intended change 
patterns were extracted by wrapping the noise and deviation 
inside irregular variation. Time series analysis was applied 
using a method which calculated base fluctuations by adapting 
the Maximum Entropy Method (MEM) as a spectral analysis 
method, and Nonlinear Least Mean Square (NLMS) algorithms 
as a temporal analysis method (Saito er a/., 1994). This method 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Figure 3. Effect of time series analysis 
539