Seismic Coherence for Discontinuity Interpretation

Abstract

Seismic coherence is of the essence for seismic interpretation as it highlights seismic discontinuity features caused by the deposition process, reservoir boundaries, tectonic movements, etc. Since its appearance in 1995, seismic coherence has become one of the most popular and highly recognized interpretation tools. In the last 25 years, there have been different kinds of approaches to calculate seismic coherence attributes, such as cross-correlation, semblance, and eigen-structure. We have also seen different coherence enhancement techniques and modified ways for coherence calculation to address different problems in diverse field applications. In this survey, we provide a general overview of seismic coherence, which is commonly used to delineate structural and stratigraphic discontinuities. We cover the development of the seismic coherence attributes via introducing existing approaches for coherence calculation, enhancement, spectral band-limited coherence methods, and offset-/azimuth-limited coherence methods. In addition, we also discuss the possible coherence artifacts and pitfalls. To better compare different techniques, field applications are provided, where different disciplinary coherence attributes with enhancement techniques are calculated for the channel, fault, carbonate karst, and volcano interpretation.

Appendix A: Artificial Intelligence (AI)-based Seismic Discontinuity Interpretation

Classic coherence calculations are typically based on the mathematical descriptions of the geophysical data representations of various structures, which can be viewed as “model-driven” methods. Seismic discontinuity detection and characterization is time-consuming, as the discontinuities need to be handpicked from raw seismic images, which may take weeks to months on a typical size seismic volume by an experienced interpreter (Bahorich and Farmer 1995). The manual discontinuity picking results are highly dependent on knowledge, experience, and even the mode of interpreters, and so human bias is inevitable.

Recently, artificial intelligence (AI) techniques have been widely applied in various fields, including audio processing, computer vision, natural language processing (LeCun et al. 2015). Therefore, there is also an increasing interest in applying machine learning and deep learning technologies to seismic data processing and interpretation (Qi et al. 2016b; Zhao et al. 2016, 2017, 2018; Wang et al. 2018b; Xiong et al. 2018; Wu et al. 2019a, b, c, 2020a, b; Wang et al. 2021; Li et al. 2021b). Machine learning (especially deep learning) technologies are powerful for mining features or relationships from data, which makes them quite suitable for learning from human experience (LeCun et al. 2015; Schmidhuber 2015). Many novel data-driven technologies have been developed in the field of deep learning to perform big data analytics automatically. One of the most popular deep learning technologies is the convolutional neural network (CNN), with successful applications in image recognition and classification (LeCun et al. 1995). State-of-the-art CNNs can recognize images even better than humans. Deep learning technologies are showing great potentials and promising results in exploration geophysics (Wu et al. 2019a, b, c; Li et al. 2021b). Thus, deep learning can be employed to assist interpreters for this labor-intensive task, such as 3D fault segmentation (Wu et al. 2019c), and automatic fault extraction based on CNN (Xiong et al. 2018). Without using “expert knowledge” or using very limited expert knowledge, AI assists the seismic discontinuity characterization via training the data-driven models based on manually interpreted labels. Xiong et al. (2018) developed a convolutional neural network (CNN)-based method to automatically detect and map fault zones using 3D seismic images in a similar fashion to the way done by interpreters. CNN methods have also been introduced to detect faults by pixel-wise fault classification (fault or non-fault) with multiple seismic attributes (Huang et al. 2017; Zhao et al. 2018). In addition, Wu et al. (2019c) not only predicted the fault probability but also estimate the fault orientations at the same time.

For the traditional seismic coherence techniques described in Sect. 2, discontinuity/coherence attributes are calculated according to the mathematical expressions of the geological structures. For the AI-based methods, data-driven models describing complex hidden features between geophysical data and interpretation are trained based on the human interpreted labels. The CNN-based fault extraction method outperforms in detecting high-angle dipping faults, while traditional image processing-based methods exhibit promising performances in vertical normal fault illustration (Qi et al. 2020). However, one typical limitation of applying deep learning models in seismic interpretation is the preparation of adequate training data sets and especially the corresponding geologic labels. To address this incomplete or inaccurate labeling issue, automatic structural interpretation can be improved by using a workflow to automatically build diverse structure models with realistic folding and faulting features (Wu et al. 2020a). AI has shown promising results in the seismic discontinuity interpretation, which is a beneficial complement to the traditional coherence calculation approaches. However, an optimal solution to this data and labeling challenge is still under development.