Texture is one of the least understood areas in computer vision. One of the major short-comings of texture segmentation approaches has been the ad-hoc selection of the set of feature vectors. We present an approach to qualitatively select a sub-set of a large (in principle infinite) set of co-occurrence matrices. A transportation measure is used to determine the difference between co-occurrence matrices resulting from various textures. This results in an ordered set of matrices, of which the resulting segmentation performance is directly related to the transportation measure. By combining segmentation results from various matrices the overall performance improves only when the matrices enhance different image areas. The most probable candidates for this can be obtained by using the same transportation measure applied to dimensional co-occurrence data. Again, this results in an ordered set. Texture segmentation results indicate a monotone increase in performance when adding subsequent matrices results from the ordered set.