Evaluation of principal component analysis for seismic attribute selection and self-organizing maps for seismic facies discrimination in the presence of gas hydrates

Mapping the lateral distribution and thickness of gas hydrates and associated free gas using even the best quality seismic surveys can be quite challenging. As the stability of gas hydrates within the Gas Hydrate Stability Zone (GHSZ) is controlled by temperature and pressure, and because their impedance can be different than those of the sedimentary matrix in which they occur, we often are able to identify a bottom simulating reflector (BSR) at the base of the GHSZ cutting across stratigraphic reflectors. However, when the stratigraphic reflectors are parallel to the ocean bottom, the BSR is easily confounded with a strong stratigraphic reflector. Seismic attributes and machine learning (ML) applied to facies analysis have been successful in quantitatively mapping patterns in seismic data. However, selecting an optimal combination of seismic attributes and training data to perform a robust seismic facies classification is highly dependent on the interpreter's skills. In this study, we apply principal component analysis (PCA) to select a suite of seismic attributes to be used as input for self-organizing maps (SOM) to identify BSRs and surrounding facies in a seismic volume located in the Blake Ridge, offshore South Carolina, USA. As PCA is based on mathematically variability unrelated to geology, we evaluate PCA as an attribute selection technique under three different training data selection strategies, finding that combining seismic attributes such as total energy, coherence, GLCM entropy, and peak magnitude as input for SOM represents an effective method to detect BSRs. Additionally, we note that using a smaller subset of samples representing the target four seismic facies when applying PCA to select input attributes for SOM offers optimal BSRs identification.