004 
5) 
the 
age 
the 
tect 
ked 
age 
XD 
int, 
zed 
the 
tial 
of 
ing 
] is 
ted 
gy 
"he 
the 
rks 
cal 
the 
ete 
ote 
the 
the 
| to 
me 
ote 
the 
1d, 
the 
um 
ect 
ing 
ark 
D 
the 
ng 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004 
space FF, = { Weert: IDI x Hs a linear 
sub-space, the watermark meeting the conditions is easy to 
search. In this paper, we only used the simple randomly seeking 
algorithm. Given embedding meaningful watermarks, a more 
complicated optimized method should be exploited, such as 
inheritance algorithm. Then we organized the wavelet 
coefficients of the sub-image into wavelet blocks, classified 
those wavelet blocks according to texture intension, by the 
classification results and according to formula (1) adaptively 
embedded the local watermark weights into different weights of 
the feature vector with different strength to obtain the new 
feature collection D -( d k) and then made contrary DWT to 
obtain the watermarked remote sensing sub-image. During the 
experiment, scale gene is supposed as a-0.25. Making the 
above process to all sub-images corresponding with all feature 
points, we can get the entire remote sensing image which was 
embedded the content-based watermarks. Figure 1 presents the 
original remote sensing image; Figure 2 is the watermarked 
remote sensing image, with PSNR=45.012dB. 
PIRES RIN sel 
SERBS IDS LN) Ghee Mtn WI m C LU s 
T QW ver Wer qum Ran TR A FIWHE NAAZ RB 
aguWée AVAL ^22 
ML dd E , est 
      
  
Figure 2. Watermarked remote 
sensing image (PSNR = 45.012) 
Figure 1. Original remote 
sensing image 
Then we emulated 11 kinds of distortion situations of 87 remote 
sensing images of different resolutions with Stirmark and made 
watermark detection by the Neymann-Pearson criterion. To 
verify the robustness and validity of the algorithm, we also 
exploited Stirmark to make robustness tests for COX algorithm 
and CBWM watermarking algorithm based on DCT (Discrete 
Cosine Transform) (Cox I J, 1997b; Ruizhen Liu, 2001a) and 
the experimental results are listed in Form 1. 
From Table 1, we can see such attacks as median filter, center 
cropping, geometrical transform, remote sensing image rotation 
and cropping in x-y directions all result in invalidation of COX 
algorithm; and CBWM entire watermarking algorithm based on 
DCT can not correctly detect watermarks only under the two 
anamorphic conditions of center cropping and remote sensing 
image rotation and can not detect watermarks when cropping of 
a remote sensing image exceeds 10% or rotation exceeds 13° ; 
however CBWM localized adaptive watermarking algorithm 
based on DWT has a strong ability against cropping. 
Experiments show only when rotation of a remote sensing 
image exceeds 15 * , CBWM localized adaptive watermarking 
algorithm based on DWT can't detect watermarks, and 
moreover the algorithm embeds watermarks into many local 
areas of the entire image, so under StirMark attack, the remote 
sensing sub-images move a little, but have little entire changes 
(MSE), therefore the entire image has a good robustness against 
StirMark attack. In addition, the total correct rate of CBWM 
localized adaptive watermarking algorithm based on DWT is 
195 
78/87=89.66 % , respectively higher than CBWM entire 
watermarking algorithm based on DCT 69/87=79.31% and 
COX algorithm 59/87=67.81%. Furthermore while making 
experiments to the remote sensing images of various 
resolutions, we found that the higher resolution, the more 
complicated texture information had the remote sensing image, 
the more robust was the watermarks in it against various 
attacks. 
We made experiments on such applications as edge detection 
(canny operator) and image classification to the original remote 
sensing image and the watermarked one respectively and the 
results has shown in Fig.3-4. During the course of image 
classification, the number of the mistakably classified pixels in 
the watermarked remote sensing image wasl025 and the 
percentage of the mistakably classified pixels of the remote 
sensing image was 0.3207%. 
E E 4 En aus 
N LA U H 
Po 
E^ 
CE 
[^ 
SOC 
   
a) Original remote sensing image — b) Watermarked image 
Figure 3. .Results of edge detection 
  
b) Watermarked image 
a) Original remote sensing image 
Figure 4. Results of classification 
From the experimental results we can see the CBWM localized 
adaptive watermarking algorithm based on DWT proposed in 
this paper has almost no influence on edge detection and 
classification of remote sensing images and is a practical 
watermarking algorithm fit for remote sensing images.