The conventional approaches for automatic seismic facies identification are based on waveform patterns using the vector information from the post-stacked trace. On the other hand, the quantitative seismic interpretation based on prestack seismic data, such as amplitude variation with offset (AVO) analysis, may bring information about lithology and fluid in the porous media. The methodologies employed for extracting signal patterns in pre-stacked seismic data generally use the combination of a post-stacking technique to extract waveform-based features with dimensionality reduction techniques. The dimensionality reduction is in fact the key point to avoid the “curse of dimensionality”. Recently, it has been proposed to combine deep autoencoders with clustering algorithms to extract seismic facies from pre-stack seismic data. In this paper, we extend this strategy, using as input, not only a single channel (narrow azimuth seismic gather), but a multi-channel data. So, we consider a set of multi-azimuthal pre-stack seismic data and frequency-gathers extracted from the post-stack data for each azimuth. Also, we use the Student’s tdistribution to build a probability map for each cluster (facies). This new strategy with the joint input is applied to a Brazilian pre-salt reservoir, showing an improvement in the recognition of the architectural elements, when compared with a single narrow azimuth input or classical approaches.

This content is only available via PDF.
You can access this article if you purchase or spend a download.