A Generalized HMM Model for Salt Dome Detection from Seismic Surveys

The oil and gas industry uses the characteristics received from the earth sub-surfaces to detect oil reservoirs. Accurate localization of oil fields is crucial to the exploration process as it can save the industry from huge financial losses which otherwise can happen due to drilling at wrong or inaccurate locations. Seismic data, obtained through the reflection of seismic waves from the earth's subsurface, contains important geological information which is used to identify a number of characteristics of earth layers such as salt domes, faults, horizons, etc. Salt domes are excellent indicators of the presence of important reservoirs such as oil and gas etc. Salt domes are mushroom shaped geologic structures that have the capabilities to trap oil and gas around them. Manual picking of salt domes is a time-consuming task due to the large size of seismic data. The accuracy of detection is also linked with the expertise of human interpreter. It is, therefore, important to develop automated salt dome detection methods which require no (or very little) input from the human interpreter. The computer-aided workflow will expedite the exploration process, increase detection accuracy, and reduces costs. In this project, we aim to develop a robust salt dome detection and tracking approach based on the Hidden Markov Model (HMM). The HMM use the concept of hidden states, therefore, it can effectively be used in seismic applications for salt dome detection and tracking. What is observed (or estimated) are the different attributes not the states themselves. For salt dome detection, we propose to use a two-state model, salt and non-salt boundary, and the texture-based features to accurately detect the salt boundaries. The optimal parameters are obtained using the training features and the backward-forward algorithm (EM algorithm). The Viterbi algorithm is used to compute the hidden states which are then used to delineate the salt boundaries. Note that we are using here only a single HMM, as such the classical classification stage is altogether avoided making the overall approach computationally very efficient. The proposed approach is designed to work with both discrete and continuous attributes.