The mechanical and transport properties of reservoir rocks depend on the morphology of microstructure, e.g. connectivity, size and shape of grains and of the pores. Such information can be gained from digital core analysis, which is increasingly used to understand the internal fabric of heterogeneous rocks or for the analysis of samples not amenable to standard laboratory analysis. At the same time, big advances are made on the imaging hardware side including the development of ultra-high resolution CCDs and recording techniques like helical scanning, leading to datasets of enormous dimensions and relatively large field of view. In this context it is highly desirable to develop automatic coarse scale classification methods to e.g. recognize the occurrence and spatial structure of digital rock types within such tomographic images - or existing morphological trends within a rock type, as this may lead to powerful characterization and data reduction techniques as well as upscaling methods.
We use regional Minkowski measures to define fine-sale rock types using a multi-variate Gaussian mixture model for classification. The discriminative power of the method is firstly demonstrated for an artificial sample which consists of a mixture of Poisson processes spatially separated using a Gaussian random field approach. Furthermore, we demonstrate how this method can be used to describe the fractions of two spatially overlapping non-stationary process generating a morphological trend - e.g. a fining up sequence. Finally, the method is applied to discriminate different morphological regimes of a thin- bedded sandstone. Importantly, for morphologies resulting from a Poisson process of grains, the classification result can directly be used to predict physical properties using effective grain shapes. For other processes such a relationship may be developed; in particular, using the classification result subsections of a tomogram can be selected for which such a relationship can be derived explicitly.