vith 
ıter, 
the 
tion 
lide 
note 
ir of 
the 
sing 
By 
for 
data 
| the 
d by 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
Take the parcels in T1 data as the unit, then select the land 
cover sample data class by class. After calculating the 
statistical data, the database of each land cover class can be 
established. When T1 data is superimposed precisely on T2 
data, take the parcels in T1 data as the unit and its land cover 
class as the reference, then calculate the corresponding feature 
statistical values in T2 data. After comparing it with the class 
knowledge information class by class in T1 data, the changed 
and unchanged positions and the regions can be detected, and 
the changed class can be recognized by matching the 
knowledge information of each class in the knowledge 
database. 
3. KEY METHODS 
3.1 Construction of the Knowledge Database of Remote 
Sensing Information of Land cover Classes 
When T1 data is superimposed on T2 data, the layers of each 
land use and land cove class can be established. In each layer, 
take the parcels in T1 data as the unit, then select the relevant 
land cover sample data in T2 image. The knowledge database 
of remote sensing information of land cover classes can be 
constructed by calculating the feature statistics of each land 
use class of the sample data. Generally, the feature statistics 
include the following values: 
(1) Spectral features, such as the spectral value of each 
band and the spectrum character curve etc. 
(2) Statistical values, such as the maximum values, the 
minimum value, mean value, variance and covariance 
etc. 
(3) Histogram features, such as the distribution, mean 
value, variance, skewness and entropy etc. 
(4) Texture features, such as self-correlation coefficient, 
entropy, homogeneity and dissimilarity etc. 
(5) Band algebra operation, such as ratio and vegetation 
indices etc. 
3.2 Construction of the Discrimination Rules 
Discrimination rules are the rulers to measure the changes of 
the land cover classes. Several discrimination rules, such as 
the Minimum Distance Rule, the Bayesian Rule and the 
Decision Tree etc. can be established according to the 
1161 
information in the knowledge database and the real-timely 
calculated feature statistics during the change detection 
process. 
3.3 Automatic Change Detection 
First, overlap T1 data and T2 data. Second, guided by the 
parcel boundary and its class information in T1 data, take the 
integrated parcels in T1 data as the unit, then compute the 
feature statistics class by class in T2 data. According to the 
given discrimination rule, and *omparing the computed value 
with the feature value of the parcels in Tl data in the 
knowledge database, the changed regions can be detected 
automatically. 
For example, T2 is the color image with R,G and B bands. 
Parcels in T1 data are used as the computing unit. Mean value 
and the variance value of each classes are used as the image 
feature value, and the Minimum Distance Rule is used as the 
discrimination rule. Then, there exists the following equation: 
  
  
  
  
  
  
T; 2 T 2 T. 2 
D. = JR (Hr Tg ) + welll — Ha V + wy (Ht, — Hp, ) 
2 = 2 NA 2 
D, ER Tr; Y Wg(O, - Ju Ex W, (0, — Oy) 
Wo We w : A 
Where, '"&, 7*6. Vg te the weight of R,GB bands 
respectively: £4 and © are the mean value and the variance 
value of each class in each band in the knowledge database: 
u 
and O are the mean value and the variance value of the 
parcel to be detected in 
PAYER SERES 
BEL LAT SD may ow see are 
* 6 Nimm iX im 
    
  
  
ua 
VIA VA LI NES SNS DIE COTE PLAN MAR OTRAS | Tissus 
WDR Kl 
45.4 AU E» ex 
Fig. 1. Change Detection Result with the Minimum 
Distance Rule 
Ls C; . 
each band. When  "" ang " are less than the given 
threshold, the parcel in T2 data is not changed, otherwise it is