Aspects of automated seismic interpretation using supervised and unsupervised machine learning

Summary The state-of-the-art seismic interpretation workflow is based on extraction of information from seismic images, which typically involves manual mapping of seed points along targeted geological structures. This process requires expert geophysical knowledge, interpretive experience, intuition and creativity. With increasing computational power, data science is continuously evolving and provide new digital tools applicable to various disciplines, including geoscience. Most of these tools are based on open-source signal processing, image processing and machine learning algorithms. By utilizing these digital tools and automate the extraction of information from seismic images, we can accumulate knowledge and build a subsurface understanding faster and better. Here, we introduce data-driven methods based on both supervised and unsupervised machine learning to address key aspects of an automated seismic interpretation workflow. We automatically identify and extract faults using a pre-trained conditional generative adaptive network together with image processing such as morphological operations. Further, we address stratigraphic units with a new 3D texture descriptor for seismic data, and compute and cluster feature vectors that describe seismic stratigraphy for given seismic sub-volumes. Finally, we correlate dislocated and truncated seismic horizons we introduce a non-local trace matching approach.