Original data set Calibration to reflectance Radiometric feature : Data fusion based on physical properties Linearity feature : Geometric context- dependant fuzzy fusion Texture feature : Geometric and parameters fusion Filtering Fuzzification Fuzzification Fuzzification Fuzzy fusions of methods and morphological post-processing : context-dependant decisions fusion Decision Segmented image Fig. 1 – Global scheme of segmentation strategy Segmentation of Hedges on CASI Hyperspectral Images by Data Fusion from Texture, Spectral and Shape Analysis M. Lennon 1 , M.C. Mouchot 1 , G. Mercier 1 , L. Hubert-Moy 2 1 Ecole Nationale Supérieure des Télécommunications de Bretagne - Département ITI Technopôle de Brest Iroise - BP 832 - 29285 Brest Cédex - France 2 Laboratoire Costel - Université de Rennes 2 - 6, avenue Gaston Berger - 35043 Rennes Cédex- France Tél : (33) (0)298001069 ; Fax : (33) (0)298001098 ; Email : marc.lennon@enst-bretagne.fr Abstract – The study figures out the potential of CASI airborne hyperspectral imagery for the fine segmentation and characterization of small size landscape units, the hedges, essential for hydrologists and landscape planners. The segmentation strategy consists in computing every hedge discriminating feature : radiometry, texture and linear shape. Original methods taking into consideration the full spectral information are developed for filtering images and computing linear and texture features. Concepts of fuzzy fusion are used to merge these information in order to get the final segmented image. Classification of the segmented region provides the bocage composition map. With the help of a DEM, 8 parameters are computed, providing a fine characterization for each pixel of the bocage. INTRODUCTION The characterization of hedges presents a great interest in the study of the diffuse pollution from the agricultural activity because they are able to absorb an important part of the water flux charged with pollution particules (nitrates, salts…) depending on their location, morphology, composition, direction [1]. Conventional techniques of remote sensing are limited for such a study : spatial resolution of satellital sensors is too low and spectral information from aerial photography are too poor for an accurate characterization. We suggest to take advantage of airborne imaging spectrometry allowing to get very fine spatial and spectral features of the hedges. The main drawback is the large amount of information to process to get the interesting thematic features. The paper shows up different levels of information processing and fusion, allowing to extract the hedges from images acquired with the CASI [2]. Images were acquired in july 1998 above Plounérin (France), region of bocage largely diseased by agricultural pollution. The spatial resolution is 2 m at ground with 9 spectral bands ranging from 400 to 900 nm. 5 flight lines were acquired, corrected from the plane attitude and mosaïcked. SEGMENTATION STRATEGY A strategy based only on the spectral features of the bocage would meet two main problems : on one hand, spectral features are multiples depending on the composition of the bocage and on the other hand, they can be the same as other landscape elements such as fields for example. Hence, other discriminant features need to be determined. First of them is the texture. If the spectral features of a group of bocage pixels is almost the same as a group of field pixels, their texture enable most often to discriminate them. Eventually, bocage is most often composed of linear structures. A shape attribute needs also to be extracted from the images. Because of the fuzzy nature of the problem and in order to homogenize the three different concepts (radiometry, texture, shape), methods of fuzzy fusion are used. Membership to each one of the three attributes is computed in parallel. Initially, original radiance images should be calibrated to reflectance in order to allow the method to be reproductible. Because of the large variability in the spectral features of original data, a filtering step is operated before the computing of shape features. This step leads to a largely less noisy result. On the other hand, the computing of the texture features must be operated on the original data set to take advantage of this variability. The result of this parallel process is a membership image to each of the 3 attributes. Concepts of fuzzy fusion are used to merge them in a single image. In order to remove the noise from this image and to take a final decision, morphological post- processing using decisions fusion depending on the context is implemented. The complete scheme of segmentation is shown on Fig.1.