Effective seismic interpretation requires understanding geologic boundary-defining attributes, which we can think of as fence posts, and lithology or reservoir-characterizing attributes, or the fence boards. With this analogy in mind, we explore how the use or non-use of these two categories of attributes can affect the interpretability of deepwater channel features in dimension reduction techniques. By evaluating traditional and machine learning interpretation approaches, we demonstrate how the preservation of both the distinct boundaries (fence posts) and lithological variations (fences) can be impacted in multi-attribute ML by the interpreter’s choice of seismic attributes.
Seismic attribute selection is a critical first step for the effective geologic interpretation of seismic data. Furthermore, the ability to align appropriate attributes with specific geologic objectives is fundamental in the identification of meaningful geologic features. As each attribute represents a calculation derived from the original seismic data, it is critical to understand both its mathematical foundation and the targeted feature it was designed to enhance. Recognition of the inherent uncertainty of each attribute underscores the interpreter’s role in verifying the use of the attribute for the interpretations. To implement this understanding methodically, one should first consider the hierarchical categories of seismic attributes (figure 1a) and select those whose application most closely meets the interpretive goal.
To demonstrate this thought process, we will show an example for channel facies discrimination in the deepwater channel system within the 3-D Romney seismic volume, in the Taranaki Basin, west of New Zealand. Based on our objectives for channel identification, we identify geometrical, instantaneous and multi-trace seismic attributes as optimal for our study goals. Once the categories are selected, we chose seven initial attributes: GLCM dissimilarity (GLCM-D), sweetness, envelope, cosine of instantaneous phase (CSP), structural curvedness (K-curvedness), most positive- and negative-curvature (K1 and K2) (figure 1b). The definitions and applications of these attributes can be found in the November 2007, December 2007, April 2014, November 2014 and December 2015 Geophysical Corner installments. After a preliminary analysis of these attributes on our data, we further narrowed them down to six attributes, opting to retain most-positive and most-negative curvature attributes in lieu of structural curvedness. Before moving forward with any ML workflows, it is important to examine each attribute and conduct an initial interpretation of internal channel architecture to verify the efficacy of the chosen attributes for our objective (figure 1c).
On the Usefulness of CSP Here
As each attribute should contribute positively to the end interpretation goal, which is channel facies identification, some discussion was needed among the research team regarding the usefulness of CSP in this project. So, let’s discuss this further. CSP is an instantaneous attribute that focuses on the phase of the seismic trace at each individual sample and is independent of the amplitude. This attribute is often used to provide a detailed view of the continuity and/or discontinuity of geological features and is often used for seismic stratigraphic analysis, as it is useful for identifying pinch-outs and terminations, and other stratigraphic geometries.
Using the analogy of fences and fence posts can help illustrate the utility of CSP in seismic data interpretation. If one imagines the sedimentary layers as shown in the seismic data as a long fence stretching across space, the CSP attribute represents the individual fence posts. Similarly, as each fence post signifies a specific point along the fence, marking boundaries and supporting the structure, the CSP pinpoints where significant phase changes in the seismic signal occur. These points of change can indicate either continuous geological features, akin to the uninterrupted stretch of a fence, or discontinuities, like breaks or gaps in the fence.
CSP versus Variance
Let us quickly consider another example of a fence-post attribute: variance. When compared to CSP for the identification of channel features, its performance is sub-par (figure 2a and b). The difference in highlighting those sharp channel edges is an important reminder to ensure visual inspection of potential seismic attributes before continuing with your workflow. In figure 2a, CSP illustrates distinct changes in the lateral reflectors and clearly delineates the unconformable surface of the channel complex edge. Within the channel of interest, we can further begin to delineate lateral channel migration and younger portions as well. The sweetness and envelope attribute displayed in figures 2d and e, as fence attributes, give us more insight into how to interpret the seismic geologically. Rather than only focusing on abrupt changes, we can use the attributes to infer something about the lithologic characteristics. In figures 2c and g we change the viewing perspective for our fence-post attribute (CSP) and one fence attribute. In this new perspective, we can follow those indications of channel features onto a time slice and get a sense of what the lateral channel morphology looks like. We have also adapted our color map for the display in figure 2g by adding bump-mapping.
Bump-mapping alters the lights and shadows, essentially creating a textured look, or relief to the seismic, which helps to delineate the channel’s architecture. In figure 2f, we note that while GLCM-D might have seemed underwhelming in the cross-sectional view for helping to really tease out the intricacies of the complex channel system, we gain a better appreciation of its ability to inform the interpreter of the changes happening in the channel system by showing it both in the bench view and by adding in the texture. For example, in figure 2g note the location indicated by the black arrow. In the cross-sectional view, we note a small area with low GLCM-D between areas with much higher values. When we map that feature onto the time-slice we see that it corresponds to a smaller-scale channel feature which is filled with fine-grained material and is clear to see.
Algorithm Selection
Once we are confident in our selection of seismic attributes for our geologic objectives, we can begin evaluating ML algorithms to select those that align with our goals. For this, we use a common ML technique to serve as a baseline to validate newer techniques. For this purpose, we selected principal component analysis. PCA is a dimensionality reduction (DR) technique that is commonly applied to 3-D seismic data when analyzing more than three seismic attributes simultaneously. While this technique has been utilized in seismic workflow for some time, it has two major limitations. First, PCA assumes that the data are linear. Second, it assumes input attributes follow a Gaussian distribution, which is not the case for many seismic attributes (sweetness and envelope excepted). Additionally, its focus on reducing dimensional complexity by focusing on variance within the data can lead to overlooking more subtle patterns that may be geologically important.
To overcome these inherent limitations of traditional PCA, we evaluate additional DR techniques specifically designed to handle more complex non-linear data, which is more similar to what we find in 3-D seismic datasets. We chose kernel principal component analysis (K-PCA) and uniform manifold approximation and projection (UMAP) for this area. Figure 3 provides a comparison between these three methods.
Application
With our ML DR algorithms selected, we use them to simplify the dimensionality of the original six seismic attributes and qualitatively evaluate each model’s ability to preserve and/or enhance the interpreter’s ability to resolve both large-scale channel features and more subtle features.
Building on this initial evaluation, we specifically assess the impact of including the fence-post attribute (CSP). In figure 4, we compare the performance of PCA and UMAP both with and without the fence-post attribute (CSP). From our original six-dimensional dataset (each attribute representing one dimension), each DR model is reduced to two components and three components (each component representing one dimension), that preserve what the model identifies as the most significant patterns in our data while simultaneously reducing the overall dimensionality. By comparing the results of these reductions both with and without the CSP, we can evaluate how significant the fence-post attribute is and also how many dimensions best preserve our geologic feature of interest. The success of these comparisons ultimately depends on the interpreter’s ability to identify both larger-scale channel boundaries and nested boundaries within the complex. CSP aids in identifying stark boundary edges and excels at defining both internal channel edges and larger-scale bounding features.
We found through this comparison that including the fence-post attribute (CSP) in DR methods plays a critical role in preserving the geologic interpretability of deepwater channel features. Including CSP in both PCA and UMAP enabled clear preservation of channel boundaries while maintaining the visibility of internal architectural elements. The notable contrast between results with and without CSP clearly shows that just as a sturdy fence needs both posts and boards, effective seismic interpretation benefits from both boundary-defining and lithology-characterizing attributes working together. While phase-based attributes are sometimes overlooked in ML workflows, our results suggest that using both types of attributes can lead to improved geological interpretations. This approach is good to keep in mind when moving from traditional interpretation methods to emerging ML techniques.
Acknowledgements
We would like to acknowledge Kurt Marfurt for spurring the conversation regarding fences and fence-post attributes. Taking the time to explore and visualize the differences has been rewarding and insightful.