This paper presents an approach to grouping natural scenes into (semantically) meaningful categories. The proposed approach exploits the statistics of natural scenes to define relevant image categories. A ridgelet-based signature is used to represent images. This signature is used by a support vector classifier that is well designed to support high dimensional features, resulting in an effective recognition system. As an illustration of the potential of the approach several experiments of binary classifications (e.g. city/landscape or indoor/outdoor) are conducted on databases of natural scenes.