A multiresolution method for analysing remotely sensed images is described, in which correlation filters, based on two-point differences, are applied over a range of scales with octave separation. When applied to typical Earth backgrounds viewed from space, the measured probability distributions of filter outputs exhibit strongly non-Gaussian statistics and satisfy scaling laws which allow a representation of the imagery in terms of fractal geometry. The method may be used as a basis for image, or image-region, characterization and, using the tails of the resulting normalised distributions, for the identification of those localized image features which are most unusual; that is, have lowest relative probability.