Asking ‘Does this attribute genuinely help me see real geologic features?’ during attribute selection is a simple habit with significant interpretive consequences. We present a practical three-question evaluation framework for seismic attribute selection and demonstrate it on a deepwater channel stratigraphic problem using the Romney 3D seismic volume, Taranaki Basin, New Zealand. Four attributes, amplitude, relative acoustic impedance, cosine of instantaneous phase, and coherence, are evaluated against two fundamental objectives: delineation of channel geometry and characterisation of internal architecture. Each attribute is assessed for what it mathematically measures, what geologic feature it can successfully enhance, and whether it delivers on that design for this specific dataset and objective. No single attribute satisfies both interpretive requirements, and attribute performance varies significantly with viewing perspective. The evaluation yields a pass, situational (dependent on geologic context and viewing perspective), or does not contribute verdicts for each attribute. The framework is transferable across geologic targets and dataset types and applies equally to traditional interpretation workflows and machine learning approaches where attribute selection directly shapes model inputs and outcomes.
Section 1: Introduction — Know your tool
Many interpreters have been in this position: you open a new attribute display, it looks interesting, maybe even striking, and you add it to your workflow without ever asking whether it is actually doing useful work for your specific problem. This isn’t carelessness; it’s a natural consequence of how interpretation software is designed and how attribute workflows are typically taught. Attributes are presented as options in a menu, and the implicit message is that more options are better. The critical question, ‘Does this attribute genuinely help me see what I am trying to see?’, may not get asked explicitly. And when it doesn’t get asked, the consequences can follow you all the way through to your final interpretation.
Let’s focus on answering that question, systematically and honestly, making it the first step in any interpretation workflow, regardless of geologic target or workflow type, whether traditional or machine learning-based (ML). To demonstrate this, we use deepwater channel stratigraphy as our test case, walking through a set of commonly used seismic attributes and applying a consistent evaluative framework to each one. The goal is not to produce a definitive ranking of attributes, because the right attribute is always dataset- and objective-dependent, but rather to model a habit of critical evaluation that any interpreter can adapt to their own problem. We begin by grounding the problem in geology, building a simple evaluation framework, and then walking that framework through candidate attributes on a real dataset.
The Romney 3D seismic volume and several of the attributes discussed here were introduced in Moreno-Ward et al. (2026a), along with a workflow for improving data visualisation through spectral balancing and attribute co-rendering. We now build upon that foundation. Here, we take a more detailed look at how attribute choice itself matters, walking through each attribute in plain language, demonstrating how we can use them to begin mapping both the morphology and the internal stratigraphic patterns of this deepwater system, and being honest about what questions each attribute answers and what questions we still have.
Section 2: Grounding the problem — What are we actually trying to interpret?
When a geologist stands in front of an outcrop, they are not simply observing; they are asking a disciplined set of questions to understand what they are seeing. Where does this body end? What do the internal characteristics within this bed indicate to me? Where are the erosional surfaces, and what do they tell me about the sequence of events? These same questions persist when we move into the subsurface, without the benefit of the outcrop to put our finger on.
Ensuring that we understand our interpretive objectives and the geologic setting allows the interpreter to better select a set of attributes that more closely align with their target. For channel stratigraphy, specifically, there are two fundamentally different things an interpreter needs to resolve, and they require fundamentally different types of attributes.
The first is geometry — where does the channel body begin and end? What are the lateral boundaries of the complex, and can we trace them continuously? This is a boundary-detection problem, and it calls for attributes sensitive to discontinuities and abrupt lateral changes in reflector character. The second is internal architecture elements — what is happening inside the channel? Are there nested sub-channels? Is there evidence of lateral migration? Can we begin to distinguish coarse-fill from fine-fill? This is a lithologic characterisation problem, and it calls for attributes sensitive to changes in acoustic properties and reflector character within a laterally continuous body. For a more in-depth discussion on geomorphology and stratigraphy in a deepwater system, see Posamentier and Kolla (2003) and Posamentier and Walker (2006).
We establish these two categories here because they set up the evaluative framework that follows. Every attribute we test will be assessed against both requirements: does it help to define the boundary, does it illuminate the interior, or does it contribute to neither?
To make this concrete, we ground our evaluation in a specific dataset: the 3D Romney seismic volume in the Taranaki Basin, offshore New Zealand, which contains a well-imaged deepwater channel complex that provides a clear test case for attribute performance. Before showing any attribute displays, we orientate the reader to the geometric features to expect using a subaerial channel, acknowledging up front that a modern river system in a subaerial environment and a deepwater channel are not a direct one-to-one comparison (Figure 1). However, the large-scale geometric features are easily recognisable across both settings, and starting with something familiar makes the transition to seismic more intuitive.
Section 3: Building the evaluation framework
Figure 2 outlines the broader interpretive workflow we operate within. It moves through three main stages. 1) Defining the geology problem, which begins with establishing the geology question, target feature, interpretation goals, and visualisation approach, with the outcome being a clear problem statement and visualisation objective. 2) Selecting and optimising attributes comes next, which involves both attribute selection and visualisation optimisation steps, including colourmap choice, co-rendering, sampling rate, and spectral balancing, with the outcome being an optimised display that highlights target features. 3) Interpreting and validating geologically, which moves through stratigraphy, geomorphology, analog comparisons, and tying architectural elements to depositional processes, with the outcome being a geological understanding and defensible interpretation. Depending on the geologic target, some of the sub-steps may alter slightly.
The evaluation framework we propose sits within the second and third stages. Before selecting any attribute, the interpreter should ask three questions in sequence. First: what is this attribute mathematically measuring? Not a vague description — a precise answer. If the answer is ‘I’m not entirely sure,’ that is valuable information and a signal to either investigate further or set the attribute aside. Second: what geologic feature was this attribute designed to enhance? Most attributes were developed with a specific application in mind, and understanding that application tells you a great deal about whether it is appropriate for your objective. Third, and most critically: does it actually perform that function on this dataset for this specific objective? The first two questions can be answered in the literature, see Barnes 2016, Chopra and Marfurt 2007 and 2008 as a good place to start. The third one can only be answered by looking at your data.
This third question is where workflows can fail. The assumption — often unstated — is that if an attribute was designed to detect discontinuities and your objective involves detecting channel boundaries, which are discontinuities, then the attribute will be useful. But attribute performance is highly sensitive to data quality, frequency content, noise levels, and the specific character of the target feature on a given dataset. An attribute that performs beautifully on a high-quality shallow dataset may be nearly useless on a deeper, noisier volume. The only way to know is to look.
The output of applying this framework to our candidate attributes is a straightforward evaluation: pass, situational, or does not contribute to this objective. We want to be explicit that ‘does not contribute’ is not a criticism of the attribute — it is a statement about the match between the attribute’s design and the specific problem at hand. An attribute that fails this test for deepwater channel stratigraphy might be exactly the right choice for a fault characterisation study on a different dataset.
Section 4: Putting the framework to work — Walking through the attributes
With the framework established and the geologic target defined, we now apply it to our candidate attributes in turn. All attributes discussed are generated from the full offset, post-stack migrated 3D Romney seismic volume. We present these in a deliberate order — starting with the most familiar before moving to the more specialised — because we want to build the reader’s confidence in the evaluation process before asking them to engage with attributes they may know less well (Figure 3). The attributes evaluated here were chosen because together they help us to understand what fundamental data we have (amplitude), boundary detection (coherence and cos Φ), stratigraphic pattern mapping (cos Φ), and intra-layer characterisation (RAI). No single attribute does all of these jobs well, which is exactly why the interpreter needs to understand the strengths and weaknesses of each attribute and how they can work well together (Chopra and Marfurt, 2007).
Amplitude is the natural starting point because it is the attribute every seismic interpreter knows best. Amplitude measures the strength of the seismic reflection, which is a function of the acoustic impedance contrast between adjacent rock layers (Yilmaz, 2001). For channel stratigraphy, amplitude is often the first display an interpreter reaches for, and with good reason — sandy channel fills commonly have a distinct amplitude character relative to the surrounding mudstone, and bright amplitude anomalies can be effective guides to channel body geometry (Posamentier and Kolla, 2003; Posamentier and Walker, 2006). However, amplitude as a boundary detection may not perform well if the channel boundary includes similar lithologies on either side, leading to gradational contrast even where the geometric discontinuity is sharp. In our dataset, amplitude gives us a strong sense of the internal character of the channel fill, but is less effective at tracing the outer bounding surface of the complex (Figures 4 and 5).
Relative Acoustic Impedance shifts us from boundary detection to lithologic inference. RAI is a pseudo-impedance attribute derived from the seismic trace by integration (Hampson et al., 2001) — it approximates the relative variation in acoustic impedance within layers rather than at their boundaries, giving the interpreter a sense of relative rock and fluid properties within the channel fill rather than just at its margins (Barnes, 2016). In our dataset, RAI provides useful additional information about the internal character of the channel body that amplitude alone does not fully capture (Moreno-Ward et al., 2026b; Figures 4 and 5). It passes our test for internal characterisation but should not be expected to contribute to boundary detection.
A legitimate question when evaluating amplitude and RAI together is whether both attributes are needed as RAI in its simplest form is derived directly from the amplitude trace. However, the two attributes do have an important difference: while amplitude response focuses on the impedance contrast at layer boundaries RAI approximates the relative impedance within the layers. Or more simply stated: amplitude indicates where layer boundaries exist with enough variability to be captured, and RAI helps us to think about what changes are happening intra-layer. Evaluating these two is not necessarily redundant as they do help to provide information for two different interpretive questions. However, in a multi-attribute ML workflow, the user must identify which attributes are genuinely useful and avoid including too many input dimensions. This would be a good case for choosing one over the other based on the primary interpretive objective.
Cosine of Instantaneous Phase earns the most discussion of any attribute in this evaluation, partly because it is often underappreciated and partly because it becomes the pivotal attribute in our discussion in channel stratigraphic mapping. Cos Φ is an instantaneous attribute, meaning it is calculated sample by sample on the seismic trace (Taner et al., 1979), and critically, it is amplitude-independent — it measures the phase of the seismic signal without being influenced by how strong that signal is (Barnes, 2016). This makes it respond to reflector continuity and discontinuity in a way that amplitude cannot, because it is not biased toward bright or dim zones. In our dataset, cos Φ provides the clearest delineation of the outer bounding surface of the channel complex in cross-section — the unconformable surface that defines the base of the channel incision is sharp and continuous on the cos Φ display in a way that is genuinely more interpretable than the equivalent amplitude or coherence display (Moreno-Ward et al., 2026b; Figures 6 and 7). It also begins to resolve lateral channel migration within the complex. cos Φ is a strong boundary-detecting attribute, and its consistent underuse in multi-attribute workflows is something that should be considered (Moreno-Ward et al., 2026a and 2026b).
Coherence is our first true boundary-detection attribute. It measures the similarity between adjacent seismic traces (Bahorich and Farmer, 1995; Marfurt et al., 1998) — where reflectors are continuous and laterally consistent, coherence is high; where they are disrupted by faults, channel edges, or any other discontinuity, coherence drops (Gersztenkorn and Marfurt, 1999; Kington, 2015). For channel boundary detection, this is a natural fit, and in our dataset, coherence does a credible job of highlighting the outer margins of the channel complex, particularly in plan view on time slices. Its significant limitation is the cross-sectional view: coherence struggles to clearly delineate boundaries in vertical section, particularly where the bounding surface is not perfectly perpendicular to the viewing direction. It also tells us very little about the internal architecture of the channel — it sees the edges but not what is between them. Coherence passes our test as a boundary-detection attribute for plan-view work but is situational in cross-section and does not contribute to internal characterisation (Figures 6 and 7).
Attribute Performance Review: Throughout the evaluation of seismic attributes, the practitioner should not lose sight of the three critical thinking questions we discussed in Section 3.
First: what is this attribute mathematically measuring?
Second: what geologic feature was this attribute designed to enhance?
Third, and most critically: does it actually perform on this dataset for this specific objective?
Focusing specifically on that third question, the interpreter should assign a verdict to each attribute: pass, situational, or does not contribute individually, or in combination with other attributes. The evaluation is qualitative and interpretive and depends on what the interpreter can see in the display, which is exactly why it has to be done explicitly and deliberately rather than assumed.
One observation that emerges clearly from this evaluation in this dataset is the relationship between attribute performance and viewing perspective. Coherence is most effective in plan view for mapping lateral extent and geomorphology; however, it does not excel in cross-section for understanding internal stratigraphic response. Cos Φ and RAI perform well in both views, with cos Φ being particularly valuable in cross-section for identifying reflector terminations, onlap, and truncation patterns that help to define channel architecture. Amplitude works well in both view perspectives but lacks the interpretive precision of the more targeted attributes.
In this application, cos Φ and coherence pass as boundary-detection attributes for lateral migration, while cos Φ is particularly effective for stratigraphic mapping, and RAI does well for intra-layer characterisation. Amplitude is best understood as situational, holding value for internal character but limited as a boundary-detection tool without support from other attributes.
The practical implication is straightforward: a well-constructed attribute suite should collectively perform across viewing dimensions, not just within a single perspective. An attribute that excels in one view and struggles in another signals that it either is not the right attribute or that it needs a complementary attribute.
Section 5: The verdict and the takeaway framework
Having walked through four attributes against a consistent evaluative framework, we can summarise the results for our specific objective — but the more important conclusion is about the process rather than the outcome. In our study, cos Φ and coherence function as boundary-defining attributes while amplitude and RAI function as internal-characterisation attributes. Understanding which category each input belongs to, and making sure both categories are represented, produces more geologically complete interpretations, whether you are working traditionally or feeding attributes into an ML workflow.
The question every interpreter should ask before opening the attribute menu is simple: Does this attribute genuinely help me see what I am trying to see, for this specific geologic objective, on this specific dataset? The broader point is this: defaulting to familiar attributes without asking that question can lead to better solutions not being utilised. The question we kept returning to throughout this process was a simple one: Does this attribute affect channel stratigraphy delineation that much? Asking it consistently, on real data, for a specific objective, is what separates attributes that earned their place in the workflow from those that did not. The same test applied to a different dataset or geologic objective might yield different results. An attribute that passed our test might fail on a noisier dataset. An attribute we set aside might be the critical one for a fault characterisation problem on a different volume.
The framework presented here is transferable even when the specific conclusions are not, and it has been developed in the context of building a single best initial interpretation that can then be built upon as needed. The authors recognise the legitimate need for ensemble models and multiple plausible interpretations, and recommend this framework as a starting point for interpreters who are newer to seismic attributes or who may not have considered seismic attributes such as relative acoustic impedance and cos Φ. Where rock properties are favourable, and offset data are available, sub-stack analysis, AVO-derived attributes, and extended elastic impedance (EEI) can further reduce ambiguity in channel fill characterisation by providing lithology-tuned seismic. The Romney dataset includes a single well located approximately 8 km outside the channel of interest, making it representative of an exploratory-stage project for which those additional workflows are not feasible. The framework presented here, however, remains applicable to those same workflows with the added dimension of well control within the feature of interest and rock physics calibration.
For the reader who takes one thing from this article, we hope it is the habit of asking the question before opening the attribute menu. Attribute selection is not a preliminary checkbox — it is an interpretive decision with consequences that propagate through every workflow step that follows it, and it deserves the same critical attention as any other interpretive decision in the geoscientist’s toolkit.
References
- Bahorich, M., & Farmer, S. [1995]. 3-D seismic discontinuity for faults and stratigraphic features: The coherence cube. The Leading Edge, 14(10), 1053-1058.
- Barnes, A. E. [2016]. Handbook of poststack seismic attributes. Society of Exploration Geophysicists.
- Chopra, S., & Marfurt, K. J. [2008, November]. Seismic attributes for stratigraphic feature characterization. In SEG International Exposition and Annual Meeting (pp. SEG-2008). SEG.
- Chopra, S., & Marfurt, K. J. [2007]. Seismic attributes for prospect identification and reservoir characterization. Society of Exploration Geophysicists and European Association of Geoscientists and Engineers.
- Gersztenkorn, A., & Marfurt, K. J. [1999]. Eigenstructure-based coherence computations as an aid to 3-D structural and stratigraphic mapping. Geophysics, 64(5), 1468-1479.
- Hampson, D. P., Schuelke, J. S., & Quirein, J. A. [2001]. Use of multiattribute transforms to predict log properties from seismic data. Geophysics, 66(1), 220-236.
- Kington, J. [2015]. Semblance, coherence, and other discontinuity attributes. The Leading Edge, 34(12), 1510-1512.
- Marfurt, K. J., Kirlin, R. L., Farmer, S. L., & Bahorich, M. S. [1998]. 3-D seismic attributes using a semblance-based coherency algorithm. Geophysics, 63(4), 1150-1165.
- Moreno-Ward, A., Bedle, H., Lubo-Robles, D., and Marfurt, K. [2026a]. Beyond Amplitude Alone: A Corendering Technique for Enhanced Deepwater Channel Visualization – a Tutorial: Systematic Optimization Using Cosine Instantaneous Phase and Relative Acoustic Impedance: Interpretation, https://doi.org/10.1190/INT- 2025-1045.
- Moreno-Ward, A., Bedle, H., Lubo-Robles, D., and Marfurt, K. [2026b] From Attributes to Geologic Interpretation: A Critical Thinking Framework for Deepwater Channel Stratigraphic Analysis, Taranaki Basin (January 01, 2026). Preprint available at SSRN: https://ssrn. com/abstract=7123941 or http://dx.doi.org/10.2139/ssrn.7123941.
- Posamentier, H. W., & Walker, R. G. [2006]. Deep-water turbidites and submarine fans. Facies Models Revisited: SEPM Special Publications, V 84.
- Posamentier, H. W., & Kolla, V. [2003]. Seismic geomorphology and stratigraphy of depositional elements in deep-water settings. Journal of Sedimentary Research, 73(3), 367-388.
- Taner, M. T., Koehler, F., & Sheriff, R. E. [1979]. Complex seismic trace analysis. Geophysics, 44(6), 1041-1063.
- Yilmaz, Ö. [2001]. Seismic data analysis: Processing, inversion, and interpretation of seismic data. Society of Exploration Geophysicists.
