FUSION OF SATELLITE IMAGES BY SPECTRAL SIGNATURES METHOD N. Ouarab, Y. Smara, L. Gueciouer & F. Hasnaoui Laboratory of Image Processing and Radiation. Faculty of Electronics and Computer Sciences. Houari Boumediene University of Sciences and Technology (USTHB), BP 32 El-Alia, Bab-Ezzouar, 16111 Algiers. ALGERIA - n.ouarab@lycos.com and y.smara@lycos.com KEY WORDS: Fusion, FCM classification, Landsat ETM, medium resolution, imaging spectrometer (MeRIS). ABSTRACT: Several works of image processing identified the advantage of merging high spectral resolution images with high spatial resolution images, in order to obtain better images and to retrieve more information in several fields of research, such as earth observation research. In this context, we propose, in this paper, the improvement of the spectral resolution of ETM images of Landsat satellite using MeRIS data of ENVISAT satellite. The two characteristics of both images (spatial and spectral) are then preserved. The technique, inspired from the MMT algorithm approach, is proposed for unmixing the data of a lower resolution by its combined processing with the data of a higher-resolution in order to generate images which have 30 m spatial resolution and 15 spectral bands. For our work, because some of missing of some data about the Algerian coastal zones and problems of registration between ETM and MeRIS images, we used for this study images from La Camargue (France). For the validation of the results, we retained the evaluation criteria of different parameters. The quality of classification plays a great role in this method (it influences directly on the quality of merged image). The best characteristics of the two images (spatial and spectral) are then preserved in the resulting image. 1. INTRODUCTION Fusion of multisensor imaging data enables a synergetic interpretation of complementary information obtained by sensors of different spectral ranges (from the visible to the microwave) and/or with different number, position, and width of spectral bands (Richards, 1999) (Minghelli-Roman, 1999) (Minghelli-Roman and al., 2001). Detailed satellite investigations of the land surface require the spatial resolution of satellite imaging instruments of within a few tens of meters, since due to the land inhomogeneity larger pixels have a high probability to be composed of various classes of land objects. To avoid a significant number of 'mixed' pixels, the resolution of the instrument should be significantly better, than a typical size of homogeneous units. In this context, the Medium Resolution Imaging Spectrometer (MeRIS) sensor, launched onboard Envisat in 2002, was designed for sea color observation, with a 300-m spatial resolution, 15 programmable spectral bands, and a three-day revisit period. Three hundred meters is a high resolution for an oceanographic sensor, but it is still too rough for coastal water monitoring, where physical and biological phenomena require better spatial resolution. On the opposite, multispectral Landsat Enhanced Thematic Mapper (ETM) images offer a suitable spatial resolution, but have only four spectral bands in the visible and near-infrared spectrum, allowing poor spectral characterization. One of the possible approaches in a multisensor data environment is to use the data of higher resolution sensors/channels to analyze the composition of mixed pixels in images obtained by lower resolution sensors/channels in order to unmix them. The purpose of this paper is to present the multisensor multiresolution technique (MMT) proposed by Zhukov (Zhukov and al., 1995, 1996, 1999) and Y.H. Hu (Y.H. Hu and al., 1999). This technique is applied by Minghelli-Roman in order to combine the spectral resolution of MERIS and the spatial resolution of Landsat ETM, the main steps of the method are implemented and the results are presented for the region of La Camargue (France). A validation method is proposed based on different statistical quality criteria. 2. METHODOLOGY The method is inspired from the MMT (Multiresolution Multisensor Technique). The MMT algorithm is adapted to the practical situation in a multisensor data environment when the detailed spatial information is available only in the high- resolution HR image. This information is used to analyze composition of the lower resolution LR pixels and to unmix them. The unmixing of the LR pixels is performed consecutively in the moving window mode. In order to unmix the central LR pixel in the window, contextual information of the surrounding LR pixels is essentially used. In particular, it is assumed that the features, that are recognizable in the high resolution HR image, have the same LR signals in the central LR pixel as in the surrounding LR pixels in the window. The algorithm includes the following operations as described in figure 1: - Geometric registration of the two images - Classification of the HR (ETM) image. - Definition of class contributions to the signal of the LR (MeRIS) pixels. - Window-based unmixing of the LR-pixels. - Reconstruction of an unmixed (sharpened) image.