Abstract
Analyzing seismic data, particularly when engaging in attribute analysis and spectral decomposition is time-consuming and requires human resources involvement. Usually, interpreters manually select and blend frequencies to identify geological patterns in seismic data. To do a proper analysis, hundreds of combinations should be generated, which introduces notable human bias.
The study objectifies the automation of spectral decomposition interpretation through the application of machine learning techniques aiming to identify and classify distinct seismic facies.
The proposed algorithm includes an integrated approach of Self-Organizing Maps (SOM) and K-Means. The input data is the frequency, spectrally decomposed from seismic data and projected on a target seismic horizon. To ensure geological relevance, a broad frequency range is considered, and steps are computed using the Structural Similarity Index Measurement algorithm. The normalized data goes through the SOM algorithm and clusters similar patterns along the frequency axis. To reduce the number of clusters, SOM nodes go through the K-means algorithm. The quality of clustering is evaluated by analyzing the unified distance matrix with weights colored by the relevant cluster.
The proposed algorithm generates a single horizon containing clustered seismic patterns. These patterns are then compared to the geologist's interpretation outcome, which was based on the "best matching" spectral decomposition RGB blending. The algorithm successfully identifies and clusters mud volcanoes along with their invasion zones and chimneys into separate clusters. It also clusters channelized facies into 2-3 distinct clusters, automatically separates floodplain areas, and identifies poor data areas.
Due to the input of a wide frequency range, the algorithm goes beyond expected geological features. It enhances channel continuity and automatically identifies thinner channelized features that were not depicted on manually created maps. The algorithm's capabilities provide valuable insights and improve the overall accuracy of seismic data interpretation.
This paper introduces a novel algorithmic approach for interpreting seismic data that goes beyond traditional methods, and offers a fresh perspective on seismic data interpretation, by enabling fully automated and quick seismic facies clustering, reducing human bias, and improving feature identification. It offers practical applications in reservoir characterization and hydrocarbon exploration, as well as valuable insights for industry professionals, researchers, and decision-makers involved in seismic data interpretation and exploration activities.
