Watermarking techniques for electronic delivery of remote sensing images M. Barni , F. Bartolini , V. Cappellini , A. Garzelli , E. Magli , G. Olmo University of Siena Department of Information Engineering Via Roma 56 - Siena 53100 - ITALY Ph.: +39 0577 234621 Fax: +39 0577 233602 E-mail: (barni,garzelli)@dii.unisi.it University of Florence Dept. of Electronics and Telecommunications Via S.Marta 3, 50139 - Firenze - ITALY Ph.: +39 0577 4796385 Fax: +39 0577 494569 E-mail: (barto,cappellini)@lci.det.unifi.it Politecnico di Torino Department of Electronics Corso Duca degli Abruzzi, 24 10129 Torino - ITALY Ph.: +39 011 5644195 Fax: +39 011 5644149 E-mail: (magli,olmo)@polito.it Abstract— This paper studies the applicability of watermarking techniques to remote sensing imagery. An overview of watermark- ing is given, and the requirements, imposed by the remote sensing scenario on watermarking techniques, are discussed. As an exam- ple, the effect of watermarking on image classification is analyzed. I. I NTRODUCTION Earth observation missions have recently attracted a growing inter- est from the scientific and industrial communities, In such systems, a spaceborne platform collects scientific data and transmits them to a ground station; at the ground segment a series of image products are created, that can be made available to scientific or commercial organ- isations for exploitation. The data delivery process, usually based on CD-ROM hardcopy or on Internet distribution, provides the user with a digital version of the remote sensing data. In the same way as for mul- timedia contents, the digital format implies an inherent risk of unau- thorized copy or use of the product; on the other hand, on the user’s side, it is important to be able to verify the integrity of the received data. Therefore, two main issues commonly arise when dealing with image data security, namely authentication and copyright protection. Both problems have been largely addressed in the field of multimedia by resorting to watermarking technology, that consists in permanently embedding a mark in the original image, carrying information such as copyright ownership and user license rights. Later on, the presence or absence of the watermark can be used to prove ownership, to protect the intellectual property rights of the document creator, to discourage unauthorized copying of the protected material, or to prove the integrity of the data. This paper is concerned with the definition of the requirements im- posed by the remote sensing scenario on watermarking techniques, in the case of copyright protection. The desired functionalities of water- marking techniques are discussed, and the possible design options of a watermarking algorithm are evaluated in terms of remote sensing spe- cific issues. II. AN OVERVIEW OF WATERMARKING Image watermarking can be seen as a communication task consisting of two main steps: watermark casting, in which the watermark is trans- mitted over the channel, which the original image plays the role of, and watermark detection, in which the signal is received and extracted from the possibly corrupted image. Intentional and unintentional attacks and This work was supported by ASI (Italian Space Agency) Grant: “Watermarking techniques for authen- tication and copyright protection of remote sensing images accessible through public and private thematic networks” distortions applied to the image, further characterize and complicate the transmission channel. As to the watermark, it usually consists of a pseudo-random sequence, with uniform, binary or Gaussian distribu- tion. According to the set of features the watermark is injected into, water- marking techniques can be divided into three main categories: (i) spa- tial domain techniques directly add the watermark to pixel values; (ii) transformed domain techniques add the watermark to the coefficients of a full-frame transform (DFT, DCT, Mellin, Radon, Fresnell) of the image; and (iii) hybrid techniques (mainly using block-wise DCT, and wavelets) working in a transformed domain, but without completely losing spatial localization. Usually, transformed domain techniques exhibit a higher robustness to attacks than spatial domain techniques. Hybrid techniques (in particular wavelet based ones) try to trade off between the advantages of spatial domain techniques in the localiza- tion of the watermarking disturb, and the good resistance to attacks of transformed domain techniques. After watermark insertion, a perceptual hiding step is sometimes per- formed to make the watermark less perceivable to the eye. In remote sensing applications watermark imperceptibility looses some of its im- portance, since the unobtrusiveness of the watermark must be judged with respect to other factors such as classification or pattern recogni- tion accuracy. Among the characteristics of image watermarking algorithms, a cru- cial role is played by the way the watermark is extracted from data. In blind decoding, the decoder does not need the original image or any in- formation derived from it, to recover the watermark. Conversely, non- blind decoding refers to a situation where extraction is accomplished with the aid of the original, non-marked data. In spite of the benefits it gives in terms of robustness, non-blind decoding is not desirable in many applications, where the availability of the original data can not be granted. An important distinction can also be made between algo- rithms embedding a mark that can be read (i.e. the bits contained in the watermark can be read without knowing them in advance) and those inserting a code that can only be detected. Watermark detection is a typical binary hypothesis testing problem. Given an observation variable, a decision rule is defined to decide whether the watermark is present (hypothesis ) or not (hypothesis ). In correlation-based detection, which up to now is by far the most common approach to watermark detection, the observation vari- able is the correlation between the watermark and the host features. To decide whether the watermark is present or not, is compared to a threshold , which is usually set by minimizing the missed detection probability subject to a maximum false detection rate (Neyman-Pearson criterion). It is worth noticing that, despite its popularity, correlation- based detection does not lead to optimum performance, unless the em-