Abstract
The use of hydrocarbons is ubiquitous in modern society, from fuel to raw materials. Seismic surveys now routinely produce large, volumetric representations of the Earth's crust. Human interpretation of these surveys plays an important part in locating oil and gas reservoirs, however it is a lengthy and time consuming process. Methods that provide semi-automated aid to the interpreter are highly sought after. In this research, texture is identified as a major cue to interpretation. A local gradient density method is then employed for the first time with seismic data to provide volumetric texture analysis. Extensive experiments are undertaken to determine parameter choices that provide good separation of seismic texture classes according to the Bhattacharya distance. A framework is then proposed to highlight regions of interest in a survey with high confidence based on texture queries by an interpreter. The interpretation task of seismic facies analysis is then considered and its equivalence with segmentation is established. Since the facies units may take a range of orientations within the survey, sensitivity of the analysis to rotation is considered. As a result, new methods based on alternative gradient estimation kernels and data realignment are proposed. The feature based method with alternative kernels is shown to provide the best performance. Achieving high texture label confidence requires large local windows and is in direct conflict with the need for small windows to identify fine detail. It is shown that smaller windows may be employed to achieve finer detail at the expense of label confidence. A probabilistic relaxation scheme is then described that recovers the label confidence whilst constraining texture boundaries to be smooth at the smallest scale. Testing with synthetic data shows reductions in error rate by up to a factor of 2. Experiments with seismic data indicate that more detailed structure can be identified using this approach.