Abstract
In this thesis we investigate the applicability of image processing and pattern recognition techniques to seismic section analysis and interpretation in oil exploration. A seismic section obtained from seismic reflection prospecting is an acoustic image which displays the cross section of the subsurface structure. The data is employed by the seismic interpreters to infer the probable location of underground resources such as hydrocarbon accumulations. However, the task is labour intensive and must be performed by human experts with sound geological and geophysical knowledge and experience. The interpretation procedure is based on visual inspection, data comparison, geologic reasoning and decision making to confirm any subsurface reserves. Hence, automating the task could have a tremendous impact in terms of speeding up the interpretation task and giving more consistent interpretation results by virtue of minimizing the element of human subjective judgement. Much work is still required to develop an automatic analysis/interpretation system. Moreover, even a partial automation of the interpretation process would bring important productivity benefits. In this context, we limit the scope of investigation herein and place the emphasis on extracting geologic events from seismic images/sections based on perceived textural appearance of the data. We adopt the stance that seismic section data can be perceived as a texture image where the character of each distinct geologic event is manifest in its textural appearance. First we quantify such seismic textures using texture representation techniques suggested in the literature. Based on the representation we develop various methods for extracting important geological events from the data automatically. In this thesis, we propose two new approaches regarding the problems of extracting seismic events at the regional and local scales: Regional scale - Seismic Section Segmentation: Image regions of different textural appearance may represent distinct geologic regions. Thus by segmenting a seismic section into regions of homogeneous textural properties it should be possible to identify and delimit regions of different seismic and therefore geological character. A supervised segmentation scheme based on the Bayesian decision rule and using a multiresolution data representation is developed and is demonstrated on seismic images with good results. Local scale - Seismic Horizon Extraction: The perceived seismic texture is constituted by seismic horizons which are caused by a change in the acoustic impedance across the subsurface rock layers and are an indication of rock boundaries in the Earth's subsurface. The seismic horizon extraction procedures result in a line map of the subsurface rock boundary structure. However, the continuity of seismic horizons is invariably corrupted by noise and variation of lithology. Hence, we develop a new robust approach using a probabilistic relaxation labelling technique with a view to automatically locate the seismic horizons and preserve their continuity. The significance of the work is twofold: firstly it provides information to the interpreter which supplements or enhances visual information available in the seismic data which otherwise are less obvious and may be ignored. Secondly, it serves as a building block for more advanced automatic analysis/interpretation system.