Fault interpretation in a vertically exaggerated seismic section : evidence of conceptual model uncertainty and anchoring

15 The use of conceptual models is essential in the interpretation of reflection seismic data. It allows interpreters to make geological sense of seismic data which carries inherent uncertainty. However, conceptual models can create powerful anchors that prevent interpreters from reassessing and adapting their interpretations as part of the interpretation process, which can subsequently lead to flawed or erroneous outcomes. It is therefore critical to understand how conceptual models are generated and applied to reduce unwanted effects in interpretation results. Here we have tested how interpretation of vertically 20 exaggerated seismic data influenced the creation and adoption of the conceptual models of 160 participants in a paper-based interpretation experiment. Participants were asked to interpret a series of faults and a horizon, off-set by those faults, in a seismic section. The seismic section was randomly presented to the participants with different horizontal-vertical exaggeration (1:4 or 1:2). Statistical analysis of the results indicates that early anchoring to specific conceptual models had the most impact on interpretation outcome; with the degree of vertical exaggeration having a subdued influence. Three different conceptual 25 models were adopted by participants, constrained by initial observations of the seismic data. Interpreted fault dip angles show no evidence of other constraint (e.g. from the application of accepted fault dip models). Our results provide evidence of biases in interpretation of uncertain geological and geophysical data, including the use of heuristics to form initial conceptual models and anchoring to these models, confirming the need for increased understanding and mitigation of these biases to improve interpretation outcomes. 30 Solid Earth Discuss., https://doi.org/10.5194/se-2019-66 Manuscript under review for journal Solid Earth Discussion started: 9 May 2019 c © Author(s) 2019. CC BY 4.0 License.


Introduction
Reflection seismic data is used to image and understand the subsurface structure of the earth, across scales and tectonic settings (e.g. Park et al., 2002;Simancas et al., 2003;Martí et al., 2008).As with other geophysical methods, seismic images are indirect representations of complex changes in the physical properties of rocks in the subsurface.Seismic images therefore carry inherent uncertainty, making them subject to multiple interpretations or, in other words, non-unique solutions (Frodeman, 1995;Rankey and Mitchell, 2003;Bond et al., 2007;Saltus and Blakely, 2011).Interpreters need to apply different conceptual models, acquired during their training and past experience, in order to produce interpretations that honour the data, particularly in areas of great uncertainty (Bond et al., 2007;Bond et al., 2015).These conceptual models are therefore the basis of the interpretation, as they provide the necessary criteria to make sense of the data (Frodeman, 1995).
To deal with uncertainty, interpreters employ heuristics (or 'rules of thumb') in the process of generating the conceptual models, and that makes them subject to a broad range of cognitive biases (Kahneman et al., 1982).One of these biases is related to the capability of interpreters to adjust their interpretations from their initial ideas or conceptual models.This type of bias, called anchoring, has been identified in many decision-making processes since it was first described by Tversky and Kahneman (1974); and takes place in the seismic interpretation process.Rankey and Mitchell (2003) investigated the effect of anchoring in an interpretation experiment by asking interpreters to reassess their seismic interpretations after being provided with additional well data.Their work shows that most interpreters did not feel that their interpretations needed to change substantially, in spite of data showing changes in porosity and net-to-gross predictions that did not fit with their initial interpretations.Their results suggest that interpreters were anchored to their initial conceptual models, and that they were reluctant to change their mind in light of new data.In a different experiment, Bond et al. (2007) observed that participants asked for the geographical location of the section and suggested that interpreters could use this information to build their conceptual models, by using geographically specific knowledge of e.g. the relevant tectonic setting to anchor their interpretation.For example, an interpreter knowing a seismic section was from the North Sea may assume a conceptual model based on an extensional tectonic regime and will consciously and unconsciously look for normal faults in the seismic data.
However, if the conceptual model is wrong, e.g.there is significant inversion in the seismic section, the interpretation could be compromised.Thus, although conceptual models are needed to develop geologically sound interpretations, they can also create anchors to potentially erroneous interpretations.
The use of tectono-sedimentary conceptual models in seismic interpretation has been extensively documented in literature (e.g.Strecker et al., 1999;Nielsen et al., 2008;Alcalde et al., 2014).Understanding what elements influence conceptual model development and application in seismic interpretations is useful to better grasp how erroneous interpretations are made.
Applying the appropriate conceptual models requires assessment, by the interpreter, of objective uncertainty, such as considering errors in data processing or acquisition, and of subjective elements, such as the potential biases they bring to the interpretation from their background and experience (Bond, 2015).Alcalde et al., (2017a) argue that image presentation also has a subdued effect in the way seismic image data is perceived and interpreted.Here, we develop this theme investigating Most 2D seismic cross-sections published in literature are displayed vertically exaggerated (Stewart, 2011), and modern computer based interpretation methods generally result in the onscreen interpretation of a vertically exaggerated seismic image, due to the conflicting ratios of a 1:1 seismic image with screen dimensions (Bond, 2015).Vertical exaggeration of seismic image data creates images with apparent reflection continuity and exaggerates dips of structures and horizons.Conscious application of seismic image stretching is used in the seismic interpretation process because it helps to enhance certain aspects of the display that ease the interpretation (Stewart, 2011).It helps for instance to amplify low relief structures, that appear otherwise compressed and difficult to differentiate (Feagin, 1981;Bertram andMilton, 1996), andBrothers et al. (2009) report that vertical exaggeration helped them to delineate small changes in stratal geometry, otherwise imperceptible, in their seismic interpretation study of the Salton Sea.Vertical exaggeration can also be used to mitigate the difference between vertical and horizontal sampling, which can be considerable depending on the acquisition parameters, the impact of which is to make images appear stretched (Stewart, 2011).However, changes in appearance of seismic image data through, sub-conscious or conscious, vertical exaggeration change an interpreter's perception of an image.The change in image character is often unintentional, and can result in unwanted perceptual bias during interpretation, and subsequently lead to misinterpretations, particularly if the interpreted geological structures are complex (Stone, 1991).Vertical exaggeration can also make features, like gas escape chimneys, appear narrower than they are (Horozal et al. 2009).Black et al. (1994) noticed that vertically exaggerated seismic sections can result in gently dipping reflections being perceived as more steeply dipping; which may lead to the erroneous conclusion that migration of the seismic data is required.Similarly, Stewart (2012) investigated the impact of vertical exaggeration on fault dip and observed that structural restoration of interpretations conducted in exaggerated sections lead to unrealistic subsurface models.Thus, vertical exaggeration in seismic interpretation can have positive and negative influences on interpreter perception of the image and interpretation outcome.
Here we test the theory that the presentation of seismic image data in a vertically exaggerated format impacts the perceptions of interpreters, influencing the conceptual models they apply in their interpretation and their final interpretation outcome.We focus on analysis of fault and horizon interpretations in a clipped seismic image.Interpreters were randomly presented with different vertical exaggerations (1:2 and 1:4) of the same seismic image.Statistical analysis of fault and horizon placement, fault dip angle, fault dip direction and fault type, allow us to draw conclusions on the effect of vertical exaggeration on interpretation.

Experiment set up
The interpretation experiment consisted of a c. 15 km long clipped portion from a 2D seismic image from the Browse Basin, NW Australia (Figure 1) available on the Virtual Seismic Atlas (www.seismicatlas.org).The seismic image has been Solid Earth Discuss., https://doi.org/10.5194/se-2019-66Manuscript under review for journal Solid Earth Discussion started: 9 May 2019 c Author(s) 2019.CC BY 4.0 License.interpreted as a series of normal faults dipping to the NW (left hand-side of the section) overlain by post-tectonic sediments, These faults could potentially have been formed in the Late Carboniferous to Early Permian rifting event (Struckmeyer et al., 1998;Keep and Moss, 2000).The area has undergone different stages of reactivation since the Early Triassic, so inversion structures can also be found (Keep and Moss, 2000).
In a series of interpretation experiments, the seismic image was presented to participants with horizontal to vertical exaggeration of 1:4 (Figure 2a) or 1:2 (Figure 2b), hereafter called 1:4 and 1:2 sections.The sections were presented in twoway traveltime (TWT) and no information about the actual depth of the sections was provided.The participants were asked to "interpret the main faults crossing the section as deep as possible", as well as to add a "sedimentary horizon to mark the displacement", and were given 15-30 minutes to complete their interpretations.The experiment as presented to the participants can be found in the Supplementary Information.
The participants also completed an anonymous questionnaire designed to collect information about their background, training, knowledge and experience in structural geology and seismic interpretation.The interpretation experiment was completed by 160 students of which 61 participants (38% of the total) were undergraduate students and 99 participants (62% of the total) were postgraduate students, from different universities in the UK, France and Spain.The participants have mostly geology (72.5%) and geophysics (12.5%) backgrounds and considered themselves as having basic to good proficiency in structural geology and seismic interpretation (>93% of the participants).We focused this experiment on students only to observe the potential variability in interpretation of the same section in a group of people with similar experience and background.

Interpretation results
The two vertically exaggerated seismic images were assigned randomly to the participants: the 1:2 section was interpreted 88 times (55%) and the 1:4 section 72 times (45%).The interpretation results were digitised manually and then converted to a 1:1 vertical exaggeration (VE=1:1) for comparison; therefore, the fault dip angles presented in this work are VE=1:1.Individual examples of the interpretation results after digitisation from both the 1:2 and 1:4 sections are shown in Figure 3.
Initially, interpretations were grouped based on fault dip direction.The majority of the interpretations dipped in a single direction, either to the left or to the right.Those interpretations with faults dipping in both directions (9.4% of the total interpretations), e.g.systems of faults and their conjugates, were not included in further analyses.Most participants interpreted faults dipping to the right (56% of the total interpretations), rather than to the left (44%) (Figure 4).The relative proportion is greater in the 1:4 sections (59% to the right) compared to the 1:2 sections (53% to the right).These two groupings were identified as it was apparent that participants interpreting faults dipping to the right and those interpreting faults dipping to the left had employed two different conceptual models to the data.This resulted in four datasets with two pairs of properties (i.e.1:2-left, notified as '1:2L', 1:2-right or '1:2R', 1:4-left or '1:4L', and 1:4-right '1:4R') that were further analysed in detail.
This subdivision allows us to study if the potential differences can be attributed to the section interpreted (i.e.1:2 or 1:4), or to the conceptual model used in the interpretation.We analysed the fault type (i.e.normal or reverse) and measured the fault dip angle interpreted by the participants.The fault type results do not show significant differences between the 1:2 and 1:4 section interpretations, with 32-33% of the participants interpreting reverse faults and 67-68% interpreting normal faults (Figure 4).However, difference in fault type can be correlated to the dip-direction of the fault (Figure 5).Only one participant (3%) amongst the left-ward dipping datasets (i.e.1:2L and 1:4L) interpreted the fault motion as reverse, while the vast majority (35 participants, 97% of the total) interpreted leftwarddipping normal faults.In contrast, most right-ward dipping faults were interpreted as reverse (56%) instead of normal (44%).
This result is more pronounced in the 1:4R, with 61% of faults interpreted as reverse, compared to the 53% in the 1:2R.
The dip angle of the faults were calculated by drawing a horizontal line at the approximate mid-depth point (1.1 ms TWT) of the seismic section, with the aim of crossing the majority of the faults around their midpoint.Similar numbers of fault interpretations were made on the 1:4 section (a total 300 faults interpreted by 72 participants), and the 1:2 section (272 faults by 88 participants) (Figure 6).The fault dip angle analyses were compared across the four datasets (Figure 7).Here we observe the biggest difference between the 1:4 and 1:2 sections, with the average dip angle of faults of 24º in the right-ward dipping, reverse 1:4 section vs 19º in the 1:2 section (Figures 7c and 7d).The fault dip of the only participant interpreting left-ward dipping, reverse faults was 23º on average, halfway between the other two groups.
There are no major differences in the analysed results across student cohorts from different universities.

Conceptual model anchoring
Analysis of participants' interpretations shows that fault interpretations in the seismic image fall into three main categories (Figure 3): (1) left-ward dipping faults with right dipping horizons (Figure 3b), corresponding to normal faulting; (2) rightward dipping faults with right-dipping horizons (Figure 3c), corresponding to thrusting; and (3) right-ward dipping faults with left-dipping horizons (Figure 3d), corresponding to normal faulting.Only one interpretation showed left-ward dipping faults with left-dipping horizons and marked the motion of the faults as reverse (Figure 5).In addition, this interpretation did not show any evidence of correlating horizons across the fault and simply used arrows to mark the motion instead.The low number of interpretations of this type (one) and the difficulty in correlation suggests that interpreting left-dipping faults with reverse fault motions is largely impossible, given the reflection seismic characteristics of the data.
Irrespective of the vertical exaggeration of the seismic image interpreted, most participants interpreted faults dipping rightward instead of left-ward (Figure 4).At the same time, the majority of right-ward dipping faults (56%) were interpreted as reverse, in contrast to left-ward dipping faults, which are mostly interpreted as normal (97%) (Figure 7).We suggest that this is as a consequence of the seismic reflection characteristics of the different features that are being interpreted as faults and horizons.Faults and horizons are interpreted in three ways (Figure 3 reflections where reflection continuity is less strong.The continuity of the right-ward dipping reflections makes them a more 'certain' interpretation than the left-ward dipping fabric.When the right-ward dipping reflections are interpreted as horizons, leaving the left-dipping fabric to be interpreted as faults, this invariably leads to interpretation of faults with normal offsets due to the angular relationship between the fault and horizon interpretations and potentially due to the participants interpretation, consciously or sub-consciously, of the nature and geometries of the basin sediments above (Figure 3b).When the right-ward dipping reflections are interpreted as faults, the sedimentary packages are harder to interpret and horizon interpretations are often forced to cut reflections (Figure 3d).When participants have interpreted faults at an angle to the rightward dipping reflections, where reflection continuity is less strong, this results in steeper fault dip angles, and interpreters often interpret the right-ward dipping reflections as sedimentary packages in horsts between reverse faults (Figure 3c).In summary, from analysis of the fault and horizon interpretations of participants, three conceptual models are identified (Figure 3) that have been applied in interpretations of the data.What we do not know is how the individual participants honed onto their 'chosen' conceptual model.The participants were prompted to interpret the faults as their main task in the experiment instructions, and as a secondary element to interpret a horizon to show fault motion; so a likely sequence is that participants interpreted faults first, although we cannot be sure that this was the case.Irrespective of the exact interpretation sequence, we suggest that once participants started interpreting certain 'features' in the reflection seismic image data as faults or horizons, they became anchored to an initial conceptual model and fitted the rest of their interpretation to this model.There is no evidence in the interpretations that the participants started off on one interpretation track and then changed this to another.Consequently, we suggest that interpreters were likely anchored to their initial thoughts on the direction of dip of the faults and the rest of their interpretation is determined by this initial model, irrespective of whether later interpretative elements conform to the data (e.g.horizons cutting reflections, as seen in Figure 3d).In such cases, although initial interpretations are informed by the data, these first conceptual models are applied irrespective of whether they later conform to the data.This has been reported by Rankey and Mitchell (2003) and Torvela and Bond (2011).This suggests that initial conceptual models play a dominant role in interpretation outcome.

Fault dip variability
Although we purport that the impact of conceptual model application and anchoring to models has the greatest influence on the interpretation outcomes of this experiment, the experiment results show certain differences in fault dip direction and dip angle between the 1:2 and 1:4 vertically exaggerated section interpretations (Figures 4,6 and 7). Figure 8 shows a projection of the interpreted fault dip angles and their averages for both the 1:2 and 1:4 sections on a graph of exaggerated vs unexaggerated dip angles.The interpreted dip angles are projected onto the corresponding curves of vertical exaggeration to show the equivalent unexaggerated dip angle.The same faults interpreted in sections with differing vertical exaggeration should have the same un-exaggerated dip angle (x-axis), but a differing exaggerated dip angle (y-axis).This is the case for the average of the right-ward dipping fault interpretations (magenta circles in Figure 8).By inference, this suggests that the same features were interpreted as right-ward dipping faults in both the 1:2 and 1:4 vertically exaggerated seismic sections.In Solid Earth Discuss., https://doi.org/10.5194/se-2019-66Manuscript under review for journal Solid Earth Discussion started: 9 May 2019 c Author(s) 2019.CC BY 4.0 License.contrast, the average fault dip angle of the left-ward dipping interpretations in the 1:2 and 1:4 sections (blue circles in Figure 8) are not aligned vertically, indicating that the two cohorts, i.e. participants interpreting the 1:2 and 1:4 sections, did not interpret the same left-ward dipping features as faults.Interpretations of left-ward dipping faults show an apparent impact of vertical exaggeration on interpretation outcome, whereas the right-ward dipping fault interpretations do not.In the 1:2 section interpretations of left-ward dipping faults have higher dip angles on average than those interpreted in the 1:4 section (Figure 8), and a greater spread in fault dip angle (Figure 6e and 6f).
The observations of fault dip angle consistency suggest that those interpreting right-ward dipping faults were unaffected by vertical exaggeration.Note that the interpreted average right-ward dipping fault dip angles are low, 20-21º; when these separated into normal and reverse faults, the right-ward dipping normal faults are very low angle 14-15º (Figure 7e-f), with the reverse faults having higher average dip angles of 19-24º (Figure 7c-d), closer to an Andersonian-predicted reverse fault dip of (30º) and falling within the range of common reverse fault dips of 10º-30º.The right-ward dipping normal fault angles however do not conform to predicted Andersonian fault dips of 45-60º (Anderson, 1905;1951), that are predominant in teaching materials (Alcalde et al., 2017c).The participants did not have access to the regional seismic line, that would have provided context for such low angle normal faults, nor to the actual depth of the sections, so participants may have been expected to attempt to interpret faults with higher dip angles to conform to accepted dip models of normal faults.We see no evidence of this and interpret this observation as data and conceptual model co-confirmation acting dominantly over other reasoning (if any took place).
For the interpretations of left-ward dipping faults, the extent of the vertical exaggeration of the interpreted seismic image appears to have an impact on interpretation outcome.Analysis of fault dip angle from the left-ward dipping fault interpretations of the 1:2 seismic section show a greater range in fault dip angle (standard deviation SD=16º) and a higher average fault dip angle of 34º, compared to the 1:4 section interpretations with an average dip angle of 24º, SD=13º (Figure 6e-f), that is, a 10º higher average fault dip in the 1:2 section.If we now consider only the participants interpretations that had also interpreted a horizon showing fault motion (Figure 7a & b), the difference in fault dip angle between the 1:2 and 1:4 sections decreases to only 3º, with similar standard deviations of 14º and 13º.We suggest that the differences observed between the 1:2 and 1:4 sections are dominated more by erroneous seismic interpretations than by vertical exaggeration, with those making 'dubious' left-ward dipping fault interpretations not completing horizon interpretations.Similarly for the right-ward dipping fault interpretations normal fault dip angles are low 24º-27º, but not as low as those interpreted to the right, suggesting that the angle of dip of the fault is driven more by the seismic image data than by any effects of vertical exaggeration.
If we consider the observations described in the light of our knowledge of the perceptual impact of vertically exaggerated seismic images (e.g.Stone, 1991;Black et al. 1994;Horozal et al. 2009;Stewart, 2012), the 1:4 section should perceptually have better reflection continuity due to data compression (Stewart, 2011).The higher apparent reflection continuity in the 1:4 section could make the right-ward dipping reflections appear more dominant and the discontinuities between the sediment packages less dominant and narrower.The smaller range in dip angles for the 1:4 section compared to the 1:2 section (SD=14º vs 16º, respectively, Figure 6a when the data is split between right-ward and left-ward dipping faults (Figure 6) and also into normal and reverse faults (Figure 7), leads us to conclude that vertical exaggeration has little impact.Our interpretation of these observations is that the seismic data and conceptual model have a more dominant influence on interpretation than any perceptual bias resulting from vertical exaggeration.

Conclusions and recommendations
We have shown in an interpretation exercise by 160 participants that: 1. Conceptual models have greater dominance on the interpretation outcome than perceptual bias from interpreting vertically exaggerated seismic sections.
2. Initial conceptual models are anchored to and there is no evidence for reassessment by participants when data does not conform to their initial model.
3. When conceptual models are confirmed, at least initially, by the data, there is no evidence that accepted models, for example in fault dip, have an impact on interpretation outcome, and that variability in interpretation (e.g.fault dips in our experiment) is minimal even if it does not conform to accepted models (e.g.Andersonian dips).Instead, the data drives the interpreted fault dip, and the conceptual model and data co-confirm each other.
Our results support the conclusions of other workers (Rankey and Mitchell, 2003;Bond et al. 2007;2008) that seismic interpreters need to be aware of potential biases when interpreting seismic image data particularly in the application of conceptual models; and of the high likelihood of anchoring to initial conceptual models even when data does not confirm or conform to the model.Research has shown that awareness of biases (e.g.George et al., 2000) can help mitigate the potential impacts of bias.Thus, seismic interpreters and their employers should employ bias awareness in their interpretation workflows, and obtain multiple opinions to test a broader range of conceptual models (see Bond et al., 2008 for workflow ideas; for reasoning tests to avoid anchoring see Bond, 2015;and Macrae et al., 2016;and for the potential impact of single conceptual models on decision making see Richards et al., 2015).Research into the effectiveness of different bias awareness techniques and their impact in geological interpretation is an obvious focus for future research.
Our work does not provide evidence, in this case, to support the conclusions of Stone (1991), Black et al. (1994), Stewart (2011and 2012) that vertically exaggerated seismic sections causes perceptual bias, compared with the dominant effect of anchoring to conceptual models.We still suggest however, that multiple visualisations of the data should be made, including at a scale of 1:1 and that care should be taken when interpretations of seismic image data have been made in a vertically exaggerated form.Other experimental work (Alcalde et al. 2017b) showed that interpreters and interpretation outcomes were influenced by seismic reflection contrast and continuity, factors that can be enhanced in vertically exaggerated seismic images.
We suggest that future work should further investigate the effect of vertical exaggeration on seismic image properties and interpretation outcomes.
Solid Earth Discuss., https://doi.org/10.5194/se-2019-66Manuscript under review for journal Solid Earth Discussion started: 9 May 2019 c Author(s) 2019.CC BY 4.0 License.how presentation of vertically exaggerated seismic image data influences conceptual model application and interpretation outcome.
): (1) along left-dipping discontinuous and chaotic reflections, these align with breaks in right-ward dipping reflections that together give the appearance of a left-ward dipping 'fabric'; (2) along 'packages' of right-dipping reflections with greater continuity; and (3) at an angle to these right-dipping Solid Earth Discuss., https://doi.org/10.5194/se-2019-66Manuscript under review for journal Solid Earth Discussion started: 9 May 2019 c Author(s) 2019.CC BY 4.0 License.
, b) may be the result of this perceptual change.But the lack of consistency in this observation Solid Earth Discuss., https://doi.org/10.5194/se-2019-66Manuscript under review for journal Solid Earth Discussion started: 9 May 2019 c Author(s) 2019.CC BY 4.0 License.

Figure captions Figure 1 :
Figure captionsFigure 1: Regional seismic image from the Browse Basin (NW Australia).The black box marks the area of the section used in the interpretation experiment.Note that the vertical exaggeration of the image is high (1:8).The full section can be downloaded from the VSA website (www.vsa.org).

Figure 2 :
Figure 2: Seismic sections used in the interpretation experiment with a) 1:4 vertical exaggeration and b) 1:2 vertical exaggeration.

Figure 3 :
Figure 3: The seismic section and sketch interpretations of the three main conceptual models applied in interpretations of the seismic section by participants.a) left-dipping normal faults with right-dipping horizons; b) right-dipping reverse faults with right-dipping horizons; c) right-dipping normal faults with left-dipping horizons.

Figure 4 :
Figure 4: Statistics for the interpreted fault directions (left 'L' or right 'R'), and motions (normal 'N' or reverse 'R').The number of participants is given in brackets.Note that ambiguous interpretations (e.g.left + right-dipping fault interpretations, or no faults interpreted), corresponding to 41 interpreters (25.6% of the total), were excluded from the count.

Figure 7 :
Figure 7: Rose diagrams showing the dips of interpreted faults and their motion.Fault dips interpreted at a vertical exaggeration of: a) 1:4, left-ward dipping and normal, b) 1:2, left-ward dipping and normal, c) 1:4 right-ward dipping and reverse, d) 1:2 rightward dipping and reverse, e) 1:4 right-ward dipping and normal, f) 1:2 right-ward dipping and normal.Note that there are fewer faults presented here than in Figure 6 due to fewer participants interpreting the fault motion.

Figure 8 :
Figure 8: Graph of exaggerated and un-exaggerated dip values for all fault interpretations, showing the average fault dips for leftward and right-ward dipping faults interpreted at 1:2 and 1:4 vertical exagertion, graph from adapted from Stewart (2011).