From 3D soil Computed Tomography images to geometrical soil structures ConfØrence Africaine sur la Recherche en Informatique 2006, Cotonou, Niger. Olivier MONGA UR Geodes, Centre IRD Ile de France ; 32, avenue H. Varagnat, 93143 Bondy Cedex, France. e-mail : Olivier.Monga@bondy.ird.fr, fax : 33148025524, tel : 33148025515 June 30, 2006 Abstract Since less than one decade geoscientists started to use 3D soil tomogra- phy images in order to achieve better understanding and modeling of soil properties. In this paper we propose one of the rst approach for dening and computing stable (intrinsic) geometrical representations of structures in 3D tomography soil images. This task sets the still open problem of describing volume shapes from discrete traces without any a priori infor- mation. Our basic idea is to describe the corresponding volume using a set of patches of volume primitives (bowls, cylinders, cones...). The volume primitives representation is assumed to optimize a criterion en- suring its stability and including a characterization of its scale (trade-o/ : tting errors/number of patches). Our criterion takes also into account the preservation of topological properties of the initial shape representa- tion (number of connected components, adjacency relationships...). We propose an e¢cient computing way to optimize this criterion using Delau- nay triangulation in an initial stage and optimal region growing in a nal stage. Region growing is performed on the adjacency valuated graph rep- resenting the primitives and their adjacency relationships. Our method is applied to the modelling of porous media from 3D soil images. This new geometrical and topological representation of the pore network can be used to characterize soil properties. Keywords : 3D soil tomography images, Computed Tomography, soil images, functionnal minimization, soil properties, geometrical modeling RØsumØ Depuis une dizaine dannØes, les chercheurs en Sciences du sol ont commencØ  utiliser limagerie tomographique tridimensionnelle. Dans cet article, nous proposons une des premiLres approches permettant de 1