Abstract
Fault detection is an important, but time-consuming task in seismic data interpretation. Traditionally, seismic attributes, such as coherency (Marfurt et al., 1998) and curvature (Al-Dossary et al., 2006) are used to detect faults. Recently, machine learning methods, such as convolution neural networks (CNNs) are used to detect faults, by applying various semantic segmentation algorithms to the seismic data (Wu et al., 2019). The most used algorithm is U-Net (Ronneberger et al., 2015), which can accurately and efficiently provide probability maps of faults.
However, probabilities of faults generated by semantic segmentation algorithms are not sufficient for direct recognition of fault types and reconstruction of fault surfaces. To address this problem, we propose, for the first time, a workflow to use instance segmentation algorithm to detect different fault lines. Specifically, a modified CNN (LaneNet; Neven et al., 2018) is trained using automatically generated synthetic seismic images and corresponding labels. We then test the trained CNN using both synthetic and field collected seismic data. Results indicate that the proposed workflow is accurate and effective at detecting faults.
Seismic data interpretation for mapping the subsurface structure thickness of top and base parigi formation, Konro field, northern West Java basin
