An Adaptive Content-Based Localized Watermarking 
Algorithm 
for Remote Sensing Image 
Xianmin Wang " Zequn Guan* 
Chenhan Wu* 
? College of Remote Sensing Information Engineering, Wuhan University! No.129 Luoyu Road, 
Wuhan City, Hubei Province, P.R.China, 430079 
wangmin10291(@sina.com 
KEY WORDS: Vision, Application, Algorithms, Feature, 
ABSTRACT: 
zequne@public.wh.hb.en 
wuchenhan(@etang.com 
Image, Spot 
In this paper, we proposed a new adaptive content-based localized watermarking algorithm for remote sensing images based on DWT, 
which adaptively embeds corresponding watermarks into the local regions of the remote sensing image instead of the entire one. In 
order to improve watermarking capability against attacks in the frequency area, we selected some stable feature points in the remote 
sensing image, then used them to identify the locations where we inserted the local watermarks, namely we embedded the watermarks 
into the local sub-images centered by those feature points. When the remote sensing image is cropped , we can also detect and orient 
watermarks by the feature points in the remaining one. Moreover the algorithm took the new watermarking-embedding strategy based 
on DWT domain to embed the local watermarks, namely watermarks should be embedded into the low frequency space firstly, and the 
remains are embedded into the high frequency spaces according to their significance. And different embedding strength for 
watermarks should be applied to the low frequency space and the high ones respectively. In order to further improve robustness of 
watermarks, the embedded local watermark is designed to be orthogonal to the feature vector of the low frequency space (LL3) in 
DWT area of the corresponding sub-image, which means that watermarking casting is remote sensing image content dependent. In 
addition, the algorithm exploited the Neymann-Pearson criterion to detect watermarks. Experimental results show that the 
watermarking algorithm is robust to all kinds of attacks, especially cropping. 
1. Introduction 
The demand for remote sensing data has increased dramatically 
mainly due to the large number of applications capable to 
exploit remotely sensed data and images. Along with the 
popularization of Internet and the development of multimedia 
techniques, people can embezzle remote sensing images 
lawlessly through the Internet. But by embedding watermarks 
into remote sensing image, we can effectually prevent such 
problems as copyright violation, illegal copying, easily forging 
and so on. 
Recently watermarking technique has developed very rapidly, 
but there is still a large distance from practical application and a 
lot of practical problems not resolved (Wentong CAI, 2001a; 
Hong Heather, 2000a; Hyung-Woo Lee, 2001a). There have 
been many watermarking algorithms, most of which can be 
classified into spacial algorithms (Hua Xiansheng, 2001a; Yi 
Kaixiang, 2001a) and frequency algorithms (He Renya, 2001a; 
Kutter M., 1999b; Christine I, 1997b; Podilchuk C I, 1998a; Xia 
X, 1997b; Pei Wang, 2002a). Generally speaking, spacial 
algorithms have poorer robustness against compressing, noise 
adding and filtering, and as for frequency algorithms, it would 
be very hard to resist cropping attack if we embed watermarks 
into the frequency area of the whole image, because now we 
have only part of image, and can't get the size of the original 
remote sensing image and the position of the cropping image in 
the original remote sensing image, then it's difficult to ensure 
the position of watermarks embedding. 
In addition, in order to ensure the robustness of watermarking 
algorithms, we think watermarks should be embedded into the 
most remarkable weight of remote sensing image. So 
watermarks should be bound together with the feature collection 
of the image, and namely, the coefficient collection 
192 
chosen to accommodate watermarks can generally be seen as the 
feature vector collection of a remote sensing image (Cox I J, 
1997b). At present, in most of watermarking algorithms the 
choice of watermarks has nothing to do with the image content, 
which in fact is quite disadvantageous to robustness and security 
of an algorithm, because when the remote sensing image with 
watermarks suffers from intentional (for instance malignantly 
destroying or watermarks deleting) or unconscious attacks (for 
example remote sensing image compression, filter, scan, copy, 
noise pollution and size change), if the embedded watermarks is 
not related to the image content, attackers would easily remove 
watermarks in the case of not destroying the basic quality of the 
remote sensing image. 
Given the above two reasons, in this paper we proposed an 
adaptive content-based localized watermarking algorithm for 
remote sensing images. The algorithm utilized the relatively 
stable feature points in the remote sensing image to mark the 
position to embed watermarks, then independently and 
adaptively embedded watermarks into that local area 
corresponding with each feature point, so when there is only 
part of the image, we can also orient and detect watermarks 
without the participation of the original remote sensing image. 
Therefore this algorithm can effectively resist cropping (In this 
paper, cropping refers to detecting watermarks in the remaining 
image after cropping, and the size of the original image and the 
position of the remaining image in the original image are 
unknown). In addition, the algorithm binds the watermarks 
together with the feature of the corresponding sub-image, 
namely the algorithm is remote sensing image content-based, so 
when the watermarks are destroyed, the features of the remote 
sensing image is also destroyed, then the remote sensing image 
would count for nothing. 
And in section 2 of the paper, we presented the adaptive