Interpretation of Geophysical Data with Higher-Level Image Processing Methods

Geophysical data can be often interpreted as abstract images containingmeasurements organized on a regular grid. Therefore, image processing methods are a natural choice to handle them. For an automated interpretation of larger amounts of data, the algorithms must generalize on data of the same kind and need to be robust against noise as well as small natural fluctuations of structures and patterns. Hence, methods are needed which model the inherent, often implicit geophysical information of interest. Higher-level image processing approaches extend standard image processing methods by results from cognitivescience, perceptual psychology, and neural sciences. These algorithmscombine image processing with statistics, computer-based learning, and elaboratemathematical tools and allow the integration of background information about the application domain and human expertise into the analysis.In this thesis, three data sets from different geophysical domains areanalyzed with higher-level image processing methods against this backdrop. The objectives are the classification of rock types in resistivity images of a borehole wall for lithology reconstruction using texture features, the segmentation of microscopy thin section images and classification of identified objects as quartz grains, pore space, and anhydrite and finally, the identification of significantarchaeological structures in magnetic data. For each of these problems an analysis framework is presented where image processing algorithms are combined in a new way or enhanced by novel methods for integrating geophysical background information. Thus, an ensemble learning classification framework is discussed for rock typeclassification. Results of several classifiers, each specialized for a certain rock type using atesting data set, are combined for improving the overall classification accuracy. The reconstructed lithology of the entire borehole corresponds to a high degree to the lithology published in the literature. For analyzing thin section images, novel feature images are derived by comparingmeasured values with a theoretical model function describing the light intensity inside uniaxial crystals. These preserve important information needed for a proper segmentation based on region competition and a classification of identified objects. Furthermore, a standard segmentation procedure is extended in this work for stabilizing the detection of boundaries between quartz grains. Finally, a perceptual grouping of a point set is carried out with the tensor voting method for reconstructing significant archaeological structures in magnetic data. These points represent sources of magnetic anomalies and are identified with the continuous wavelet transform for potential fields. The tensor voting identifies points being part of larger structures which form arcs and lines indicating relevant archaeological building remains.The results worked out in this thesis for the three different applicationsindicate that the higher-level image processing approach, i.e. combining image processingwith learning, statistics, and mathematical modeling of background information, iswell suited for broader applications in geosciences