The crossdip correction as a tool to improve imaging of crooked line seismic data: A case study from the post-glacial Burträsk fault, Sweden

Understanding the development of post-glacial faults and their associated seismic activity is crucial for risk assessment in Scandinavia. However, imaging these features and their geological environment is complicated due to special challenges of their hardrock setting, such as weak impedance contrasts, sometimes high noise levels and crooked acquisition lines. A crooked line geometry can cause time shifts that seriously de-focus and deform reflections containing a crossdip component. 5 Advanced processing methods like swath 3D processing and 3D pre-stack migration can, in principle, handle the crooked line geometry, but may fail when the noise level is too high. For these cases, the effects of reflector crossdip can be compensated for by introducing a linear correction term into the standard processing flow. However, existing implementations of the crossdip correction rely on a slant stack approach which can, for some geometries, lead to a duplication of reflections. Here we present a module for the crossdip correction that avoids the reflection duplication problem by shifting the reflections prior to stacking. 10 Based on tests with synthetic data, we developed an iterative processing scheme where a sequence consisting of crossdip correction, velocity analysis and DMO correction is repeated until the stacked image converges. Using our new module to reprocess a reflection seismic profile over the post-glacial Burträsk Fault in Northern Sweden increased the image quality significantly. Strike and dip information extracted from the crossdip analysis helped to interpret a set of southeast dipping reflections as shear zones belonging to the regional scale Burträsk Shear Zone (BSZ), implying that the BSZ itself is not a 15 vertical, but a southeast dipping feature. Our results demonstrate that the crossdip correction is a highly useful alternative to more sophisticated processing methods for noisy datasets. This highlights the often underestimated potential of rather simple, but noise-tolerant methods, in processing hardrock seismic data.

. The majority of the post-glacial faults strike north-northeast or northeast but both the Burträsk and Röjnoret faults deviate significantly from this trend. Deformation zones ©Geological Survey of Sweden.

Geological setting
The survey area is situated in the Paleoproterozoic rock formations of the southern Skellefteå District. The main structural feature of the area is a wide, dextral shear zone, suggested to have formed by lateral escape during the Svecokarelian orogeny (Romer and Nisca, 1995). Following Romer and Nisca (1995), we will refer to this feature as the Burträsk Shear Zone (BSZ).
It was subdivided further by , but for simplicity we continue using the original definition.

5
The BSZ marks the transition from metasedimentary rocks in the south to an area dominated by magmatic rocks in the north (Fig. 2). The metasedimentary rocks belong to the Bothnian supergroup -a sequence of sediments of mostly turbiditic origin that was accreted and metamorphosed during the Svecokarelian orogeny. They consist mainly of highly deformed and migmatized meta-graywackes, meta-argillites and paragneisses (Kathol and Weihed, 2005). The magmatic rocks in the northern part are attributed to have originated from two different phases of magmatism: The early Svecokarelian calc-alkaline intrusive 10 rocks are mostly of granodioritc or tonalitic composition and are tentatively dated to 1. 96-1.86 Ga (Kathol and Weihed, 2005).
The late to post Svecokarelian granites likely originated from magmas derived from the middle and upper parts of the crust during a period of intense deformation and regional metamorphism at approximately 1. 82-1.76 Ga (Kathol and Weihed, 2005). However, the exact tectonic evolution and timing of the magmatism in the area is still debated and a couple of different models exist Rutland. et al., 2001;Weihed, 2003;Juhlin et al., 2002;Lahtinen et al., 2009;Skyttä et al., 2012).
Similarly, the age of the BSZ is still discussed and suggestions range from 1.825 Ga (Romer and Nisca, 1995) to 1.86 Ga Rutland. et al., 2001;Skiöld and Rutland, 2006) and 1.895 Ga (Weihed et al., 2002). However, most authors agree that the peak of regional metamorphism in the area was around 1.825 Ga Rutland. et al., 5 2001; Weihed et al., 2002) and no major re-activation has so far been documented after 1.79 Ga. Since no borehole or seismic data are available, the geometry of the BSZ at depth is interpreted from surface observations and tectonic considerations. Romer and Nisca (1995) suggest a vertical strike-slip zone whereas  prefer a south side up dip-slip system.
The Burträsk fault scarp consists of a series of mostly southwest-northeast oriented segments, together forming a c. 35 km long lineament (Fig. 2). In the easternmost part, it follows a deformation line belonging to the BSZ and in the central part, 10 it continues sub-parallel to the BSZ, cutting through a large intrusion east of Bygdeträsk (Fig. 2). In the westernmost part, the deformation zones of the BSZ change direction to a more east-west orientation and the fault scarp starts to diverge from the BSZ (Fig. 2). The fault scarp is usually 5-10 m high and generally covered by a variable layer of Quarternary sediments dominated by till, clay and silt. In a few locations, scarp outcrops forming slightly overhanging cliffs are observed (Lagerbäck and Sundh, 2008). Based on sediment liquefaction features close to Umeå, Mörner (2005) suggested a M>7 event for the 15 formation of the fault scarp. On the other hand, Lagerbäck and Sundh (2008) found several sediment liquefaction and water escape structures during an extensive study of sediment deformation in the area, but were not able to establish any relationship between the intensity of deformation and the proximity to the fault scarp. One of the main challenges of crooked acquisition lines is that the trace midpoints have an offset component perpendicular to the profile direction. In the following, we will refer to this component as 'cross-offset'. As a consequence, the depth to a reflector with a crossdip component can vary for individual traces in a CDP gather, leading to an additional term in the traveltime equation (Nedimović and West, 2003): where t is the traveltime; t 0 is the zero time of the reflection; x is the inline offset; y is the cross-offset; h is the source-receiver distance; p is the slowness in profile direction and p y the slowness perpendicular to the profile.
These time shifts are not accounted for in the standard NMO processing which can result in focusing problems. Larner et al. (1979) defined the crossdip correction ∆ tcross in the form of: where φ is the crossdip angle and v is the velocity.  The underlying assumption of this correction is that the reflector has no dip component in the profile direction (see Nedimović and West (2003) for a more general form of the crossdip correction).
The effect of uncorrected crossdip on the stacked seismic section depends mainly on the distribution of the cross-offset, i.e. on the geometry of the acquisition line. Figure 3 shows a synthetic example of an NMO corrected gather including two reflections affected by crossdip at 0.4 s and 1.2 s, respectively. The cross-offset distributions of these reflections correspond to 5 the distributions at CDP 350 and 1350 of the Burträsk profile (Fig. 2). The energy of the upper reflection (A) is completely smeared and is hardly visible in the stack. For the lower reflection (B), one cross-offset value is dominating the whole dis- tribution, causing the energy to focus at 0.8 s instead of 1.2 s. As a result, the reflection appears in the stack, but shifted in time.
Unlike the case for a reflector dipping in the profile direction, the crossdip correction has no lateral component, but it requires specific knowledge of both crossdip angle and velocity in the cross-profile direction or cross-slowness, respectively. Since these parameters are usually unknown, they have to be derived from the data.

5
Despite the simple form of Eq. 2, the correction has been applied in quite different ways by different authors. Initially, Larner et al. (1979) calculated the crossdip correction simultaneously with the residual statics solution. In their fundamental paper, Nedimović and West (2003) developed a procedure for an automated crossdip correction where they invert for the cross-slowness by a grid search using the product of semblance and a local running average of the amplitudes as the objective function. For each estimated cross-slowness, they evaluate the reliability by thresholding the stack's amplitude and modal 10 filtering. Finally, they correct for the estimated crossdip while stacking using a slant stack approach. Other authors have used the same slant stack approach, but with manually determined crossdip values (Kim and Moon, 1992;Kim et al., 2014). A more simplistic approach is applying the crossdip correction as a static shift to the whole trace (Lundberg and Juhlin, 2011;Ahmadi et al., 2014). However, this approach generally decreases the quality of all other reflections in the trace and is consequently mostly used for analyzing the crossdips without applying the correction (Urosevic et al., 2007;Rodriguez-Tablante et al., 2007;15 Malehmir et al., 2009;Dehghannejad et al., 2010Dehghannejad et al., , 2012Hedin, 2015).
While the problems with applying the crossdip correction as a static shift are obvious, there are some more subtle disadvantages with implementing it using the slant stack approach. First of all, the slant stack procedure makes any further processing after the crossdip correction impossible. This might be problematic since there are very likely interactions between crossdip, DMO and NMO corrections. Another issue occurs when a CDP gather is dominated by one cross-offset value. In this case, 20 the reflection occurs twice in the stack: at the origin time as well as at the shifted time corresponding to the dominating crossoffset (red arrows in Fig. 3c). To avoid this reflection duplication, we decided to use a method which moves the energy of a crossdipping reflection to its origin time. Furthermore, we opted for a manual analysis of the crossdip angles since automatic detection as in the method of Nedimović and West (2003) is susceptible to noise and might pick up inline dip variations and NMO residuals.

A new module for local crossdip correction
We developed a module for a local crossdip correction of individual reflections that can be directly used in a commercial software package. In our module, the reflections are approximated by a polygonal chain with the crossdip values defined at the vertices of the chain. Between the vertices, the crossdip values are linearly interpolated. For the actual correction, we use a cut-and-shift method where reflections affected by crossdip are cut in a window around the biased reflection time t 0 + ∆t cross , 30 shifted back by the correction term ∆t cross and added to the trace at their origin time t 0 (Fig. 3). This procedure can lead to gaps in the corrected trace, but is necessary to prevent reflection duplication as in the case of reflection B in Fig. 3c.
Along with the crossdip values, the user defines the width of the correction window at each vertex and a constant velocity for the whole profile. We have decided to use the crossdip angle as the main parameter since it is more intuitive than the cross-

) indicates that reflection
A is missing and reflection B is shifted. In the slant stack (c), reflection A is correctly imaged but reflection B appears twice in the trace.
After crossdip correction, the energy is aligned at the correct time in the CDP gather (d) and in the corrected stack (e), both reflections are correctly imaged.
slowness, but it is very important to be aware of the coupling between crossdip angles and velocities. Therefore, we recommend to translate variations and uncertainties in the velocity into a range of possible crossdip angles for geological interpretation.
The module also includes an interactive function for analyzing the crossdip angles. Analogous to velocity analysis, the crossdip values can be picked interactively on panels corrected with a constant crossdip angle. Thereby, it is possible to observe the effect of the correction on a whole reflection and not just on a CDP gather.

4 Synthetic tests
We created a series of synthetic test datasets using the same acquisition geometry as in the Burträsk dataset. The modeling is based on the simple modeling approach outlined by Ayarza et al. (2000). In this method, the traveltimes are calculated from the geometry of the acquisition line and the reflectors, while the amplitudes are obtained using the formulas of Aki and Richards (1980). The seismic traces are created by convolution of a scaled spike at the calculated traveltime with a Ricker wavelet.

Model 1
The first model consists of a series of reflectors with increasing crossdip in a constant velocity medium (Fig. 4a). The objectives of this model were to test the correction method and to analyze the effects of crossdip on the stacked section. Figure 5a shows a stack of the synthetic data from model 1 using the true model velocities. Depending on the geometry of the acquisition line and the dip angle, the reflector crossdip manifests itself as interference effects and smearing of the reflections in the stacked section.

5
A subsequent analysis of the optimum stacking velocities yielded alternating high-/low-velocity patches mimicking the mean cross-offset. However, the stack quality remained poor for the reflections with larger crossdips since the NMO correction can only account for hyperbolic traveltime distortions (Fig. 5b). When applying our crossdip correction module, we were able to retrieve the correct crossdip angles for all reflectors and the reflections in the stacked section were effectively focused and flattened ( Fig. 5c). The second model was set up to evaluate how much inline dip can be picked up by the crossdip correction and consisted of a series of plane reflectors with increasing inline dip in a constant velocity medium (Fig. 4b). As expected, it was not possible to find any crossdip angles that consistently improved the stack of this dataset. Applying the crossdip correction clearly distorts the reflections (Fig. 6). Some very localized focusing also occurs, but overall, the smearing of the reflections

Model 3
The aim of the third model was to test the interactions of the crossdip correction and the DMO correction and to establish the preferred order of these steps in the processing flow. The model features a series of reflectors with different ratios of inline and crossline dip in a constant velocity medium (Fig. 4c). Figure 7 shows two versions of the stacked section with (a) crossdip correction applied before DMO correction and (b) after DMO correction. In the first case, picking the optimum crossdip angle 5 proved to be a bit challenging due to local maxima caused by the uncorrected inline dip. However, these occurred within a few degrees of the correct value and we were able to retrieve the expected values when picking with a focus on consistency. In the second case, similar difficulties arose because of artifacts introduced by the DMO correction. Again, focusing on consistency helped to identify the optimum crossdip angles. In both versions of the stack, the reflections become flattened and focused, but the first version, with the crossdip correction applied before the DMO correction, contains less artifacts and is slightly more pplied. The reflections get clearly distorted due to the falsely applied crossdip correction. In most areas, smearing is increased but locally, reflections focus slightly.

Implications for the implementation of the crossdip correction
As demonstrated clearly by the results of the first model, crossdip effects can be picked up by stacking velocities, manifesting themselves as alternating high-/low-velocity patches. Therefore, it is important to re-analyze stacking velocities after correcting for crossdip and to review the obtained crossdip angles with the updated velocity model. Furthermore, the results from the second and third model highlight the importance of picking only crossdip angles that consistently improve the image of a 5 whole reflection (segment) in order to exclude local optima caused by interactions between the effects of inline and crossline dip. The advantage of the manual crossdip correction approach is that it is possible -to a certain extent -to identify and avoid these interactions whereas the automatically conducted DMO correction is inevitably influenced by them.
Based on these results, we recommend an iterative processing sequence, where the crossdip correction is applied after an initial velocity analysis, followed by repeated velocity analyses and DMO correction. As mentioned above, the crossdip angles The Burträsk profile was recorded on 280 channels with 20 m spaced 28 Hz single component geophones. The signal was generated using a VIBSIST hydraulic hammer source (Cosma and Enescu, 2001) with a nominal shot spacing of 20 m. Since the source could not be activated in the vicinity of buildings, shot coverage is quite sparse close to inhabited areas. As a result, 5 the fold varies considerably over the profile (Fig. 9a). Similarly, the data quality is affected by random noise and coherent noise originating e.g. from the source, surface waves and ground roll to a varying extent. All shot gathers have a relatively high background noise level, but some feature very distinct first arrivals and clearly visible reflections whereas others show mostly noise with first arrivals barely identifiable after a few hundred meters (Fig 8). As a first order estimate of the average signal-to-noise ratio, Fig. 9a illustrates the ratio between the mean amplitude of the background noise in a 200 ms window 10 before the first arrival and the mean amplitude of the first arrival in a 200 ms window starting at the first arrival (see also Fig   8).
The first part of the reprocessing followed the original processing flow from Juhlin and Lund (2011) quite closely. Due to the high noise level in some areas, we re-picked the first arrivals manually and re-analyzed the stacking velocities. We applied an additional ground roll and first break muting step following Oren and Nowack (2018). In their method, ground roll and first breaks are estimated by soft thresholding in the local time-frequency domain (Liu and Fomel, 2013) and subtracted from the data. Apart from that, there are only minor differences to the original processing, including slightly lower bandpass filter frequencies and a spherical divergence correction instead of an automatic gain control. A summary of the processing is given in Table 1.
Following the procedure outlined in the Sect. 4.3, we incorporated the crossdip correction directly after residual statics 5 corrections, re-analyzed the velocities, applied a DMO correction and analyzed the velocities again. With the new pre-DMO velocity model, we updated the crossdip angles and re-applied the DMO correction using the post-DMO velocity model. During the first pass of this procedure, the velocities changed substantially after each step, but converged during the second pass. The final stacking velocity model is much more consistent along individual corrections than the initial one, but seems to be still For comparison, we also tested applying the crossdip correction after the DMO correction, but could not produce a stack of comparable clarity.
Migration testing using a smoothed stacking velocity model for migration yielded poor results, confirming that the stacking velocities do not correspond to true subsurface velocities. The best results were archived for a Stolt migration with a constant velocity of 5400 ms −1 . This velocity is also consistent with the velocities obtained during the refraction statics correction.

5
Finally, the section was depth converted using the same constant velocity. However, it is important to note that the depth values are only an estimation of the real depth since considerable uncertainties exist in the deeper part, where the velocity is poorly constrained.
Additional to the reflection seismic processing, we carried out first break tomography using the PStomo_eq solver by Tryggvason et al. (2002Tryggvason et al. ( , 2009). We carried out the inversion in 3D with a cell size of (x, y, z) = (20 m, 20 m, 10 m) while simulta-10 neously estimating a statics solution to eliminate the influence of sub-grid scale velocity variations on the model.  of the scarp segment west of the seismic line around CDP 1600 (Fig. 2). At the same location, the geological map features a deformation zone, but the strike does not agree well with the estimated strike of reflection B1 (Fig. 2).

Results
Reflection B2 branches in the upper part, therefore we used two different inline dip values in the estimation.
Quaternary sediments and the bedrock, most rays were guided along the bedrock surface and did not penetrate into the deeper parts.
In the upper part of the bedrock, the velocity is mainly around 5300-5500 ms −1 and increases slightly with depth (Fig. 11).
The model features several low velocity zones; the largest one at around x = 4.0 km to x = 4.5 km and coinciding with the location of the fault scarp, some smaller ones around x = 5.0 km, x = 5.5 km and x = 9.0 km and a couple of very localized 5 ones throughout the whole profile.
6 Discussion and Interpretation

Crossdip correction
Both synthetic modeling and the field data example have clearly demonstrated the benefits of applying the crossdip correction to crooked line seismic data. As several previous studies have already illustrated, crossdip can de-focus and smear reflections, 10 resulting in a poor stacked image (e.g. Larner et al., 1979;Kim and Moon, 1992;Nedimović and West, 2003). Another quite rarely mentioned aspect is that it can lead to vertical shifts, distortion and duplication of out-of-plane reflections, as well. Our local crossdip correction addresses this problem by shifting back reflections affected by crossdip and thereby projecting them into the CDP plane. However, the drawback of this procedure is that it can not handle crossing reflections very well. Therefore, the applicability of our method to areas with complex, distributed reflectivity patterns is limited. In such areas, more extensive 15 testing is needed to develop an appropriate correction scheme since existing schemes, like the one of Nedimović and West (2003), do not account for the reflection duplication problem.
Apart from improving the imaging of crooked line seismic data, the crossdip correction has the advantage of extracting information on the 3D orientation of reflections, information which can be crucial for the geological interpretation of the seismic data. This can be especially important in low signal-to-noise data with low fold where sparse swath 3D processing is 20 not an option.

Data Reprocessing
Processing crooked line seismic data from hardrock settings includes a whole range of different challenges. Apart from crossdip effects, imaging quality is often affected by residual static shifts, surface waves, coherent noise, strong ground roll, etc. As in many other studies (e.g. Juhlin, 1995;Pretorius et al., 2003;Urosevic et al., 2007;Place and Malehmir, 2016), the most important steps in the basic reprocessing of the data proved to be the manual picking of first arrivals for improving the static 5 correction and a careful velocity analysis. Furthermore, the comparison between Fig. 9b and 9c illustrates that the final image benefited considerably from using an analytical gain instead of applying an AGC since the contrast between the main reflections and the background reflectivity is preserved.

Origin of the reflections
The reflections in the final stack are mostly planar and occur in a relatively low reflectivity surrounding (Fig. 10). There are 10 different possible origins for such reflections. The first possibility is a contact between different lithological units. Along the CDP line, the surface geology map features several lithological contacts (Fig. 2 & 10), but these contacts will only produce reflections if there is a significant difference in seismic impedance between the two lithological units. Both the output bedrock velocity from the refraction statics correction and the first break traveltime tomography do not show any significance changes in bedrock velocity along the profile. This observation along with the lack of correlation between the seismic reflections and 15 surface geology (Fig. 10) and the fact that some reflections are clearly cross-cutting sub-horizontal background reflections ( Fig. 10), make lithological contacts a rather unlikely candidate. A second possibility is deformation or shear zones with either decreased or increased seismic impedance due to fracturing and/or re-mineralization processes. Whether or not such zones are visible in seismic data depends not only on the impedance contrasts, but also on the width of the zones. Theoretically, features with a minimum width of λ/30 − λ/20 (Sheriff and Geldart, 1995), corresponding to 2.4-4.5 m for a peak frequency of 60-75 20 Hz and a velocity of 5400 ms −1 in this survey, are detectable in seismic reflection images. In practice, the detectability limit depends strongly on the signal-to-noise ratio of the data. In the Burträsk survey, the signal level is relatively low, so a more realistic estimation of the detectability limit is λ/12 − λ/8, corresponding to a width of 6-11.25 m. Since the profile intersects a couple of deformation zones belonging to the BSZ (Fig. 2), it is quite plausible that some of the reflections are caused by shear zones. Another possible scenario is that some reflections originate from magmatic dykes and/or sills. The Burträsk 25 area was subject to intense migmatization and hosts intrusions from different magmatic pulses prior to, during and after the Svecokarelian orogeny (Kathol and Weihed, 2005). Thus, both sills and dykes are likely to occur along the profile.
In the following, we will discuss the nature of the most important reflections in the Burträsk profile. B1 is a relatively weak planar reflection that can only be observed clearly over a short depth interval, but in the unmigrated stack there are some indications that it might continue to greater depths. (Fig. 9) Moreover, it seems to cut through the mostly sub-horizontal 30 background reflectivity (Fig. 10), arguing against a lithological boundary. Since the surface projection of reflection B1 coincides both with a projection of the western scarp segment (Fig. 2) and a low velocity zone in the tomography model (Fig. 11), we interpret B1 as a reflection from either the western scarp segment itself, or from the continuation of the shear zone along which the western scarp segment has ruptured. The small vertical extent of the reflection might be explained by a rather narrow shear zone that drops below the detection limit as the frequency content and signal strength decrease with depth.
Similar to B1, reflection B2 cross-cuts sub-horizontal background reflections (Fig. 10) and its surface projection coincides with a narrow low velocity zone in the tomography model (Fig. 11). Again, we interpret B2 as a reflection from a shear zone, but the relation to the Burträsk fault is less obvious. The prominence of the reflection might suggest that the movement of the 5 post-glacial fault at depth took place along reflection B2 and that the visible fault scarp segments are merely the branches of that fault where the surface rupture occurred. Branching of post-glacial fault scarps seems to be a common phenomenon and has has been observed for the Pärvie and Lansjärv faults (Talbot, 1986;Juhlin et al., 2010;Ahmadi et al., 2015). Since the Burträsk fault is seismologically very active, earthquake locations might give a hint which fault plane is active at depth. Recent studies show the micro-earthquakes clustering along a southeast dipping plane, but unfortunately, the accuracy of the locations 10 is not sufficient to distinguish between the closely spaced reflections B1 and B2 (Lund et al., 2016). In any case, B2's estimated strike of 61 • -70 • is not consistent with the strike of the fault scarp but rather matching the trend of the BSZ in the southern part ( Fig. 2). Therefore, we prefer to interpret B2 as a reflection from a local shear zone belonging to the BSZ and not connected to the Burträsk fault. It is, however, still possible that B2 extends further to the southwest and connects with the westernmost scarp segment.
Since reflection B4 projects to the surface close to a mapped deformation zone (Fig. 2), we tentatively interpret it as another local shear zone belonging to the BSZ.
Both of the two deepest reflections, B3 and C1, exhibit strong reflectivity and terminate very abruptly at the upper end.
Neither data fold and amplitude ratio (Fig. 9a), nor the overall impression of the image quality ( Fig. 9d & 10), indicate a significant drop in data quality, so the abrupt terminations seem to be real features. Together with the high reflectivity, this 20 suggests that reflections B3 and C1 are caused by sill or dike intrusions.
Unlike most of the other reflections, A2 and A3 are dipping to the northwest and their approximate surface projections are well south of the BSZ (Fig. 2). At 0.8 s, reflection A2 is clearly intersecting with a sub-horizontal reflection segment ( Fig. 9d & 10), precluding the possibility of a lithological contact. The geometry of reflections B2, A2 and A3 and the apparently lower dip of reflection A3 could be interpreted as a positive flower structure, consistent with the proposed oblique convergence of the 25 Skellefteå district from the southeast (Bergman- Weihed, 2001). However, the true dips of reflections A2 and A3 are relatively small and very similar (Table 2) and the estimated strikes of reflections A2, A3 and B2 do not match at all (Fig. 2). So in this case, the additional information from the crossdip analysis argue against the hypothesis of a flower structure. Instead, the occurrence of several magmatic bodies southeast of the profile points towards a magmatic origin of the reflections, possibly as feeder dikes following pre-existing weak zones in the upper crust. It is not clear from the present data set how reflection A1 30 should be interpreted.
Even though only reflections B1 and B4 can be directly correlated to deformation zones in the geological map, the dominance of southeast dipping reflections in the northwestern part of the profile suggests that the BSZ formed as a south-side up dip-slip system as interpreted by  and not as a vertical strike-slip zone as suggested by Romer and Nisca (1995). Another possibility is that the reflections from the fault scarp are not stacked properly due to the complex, three-dimensional geometry of the fault at the profile location. In this case, the reflections should still be clearly visible in the shot gathers. The shot gathers, however, exhibit only very blurred dipping reflections covered by various forms of noise. Figure 9a shows that between CDP 1600 and 1800, both the data fold and the average amplitude ratio between first arrival and the background noise 10 are exceptionally low, indicating a poor signal-to-noise ratio. Therefore, the most likely explanation for the lack of a reflection connected to the fault scarp is simply the result of insufficient data coverage and quality.

Relation between post-glacial fault and BSZ
The relation between the Burträsk fault and the BSZ is still not fully understood. Compared to the majority of the known postglacial faults, which predominantly strike north-northeast, the Burträsk fault has an unusually strong east-west component. In

15
contrast to this, the neighboring Röjnoret fault is mostly north-south oriented, following yet another set of Paleoproterozoic shear zones (Fig. 1). This divergence from the main trend of post-glacial faults might indicate that the faults in the Skellefte area were to a very large extent guided by pre-existing weak zones in the crust. However, the Burträsk fault only follows individual deformation zones closely in the northernmost part and runs sub-parallel to the BSZ in the central part. Since the reflection seismic image has shown that there are potentially many more shear zones than the ones marked in the geological map, it is 20 likely that the fault still follows weakness zones belonging to the BSZ. South of the large jump north of Bygdsiljum (Fig. 2), the fault scarp starts to diverge significantly from the BSZ where the latter turns to a more east-west orientation, suggesting that the orientation of the BSZ was no longer conducive to the prevailing stress field controlling the rupture direction. Therefore, we speculate that the BSZ acted as a guide for the Burträsk fault, causing its orientation to deviate from the orientation of the majority of post-glacial faults. During the rupture, the fault probably jumped between different weak zones to accommodate 25 differences between the orientation of the BSZ and the orientation of the minimum horizontal stress.

Conclusions
In the first part of this paper, we presented a new software module for a local crossdip correction and tested the influence of crossdip on synthetic seismic data. An often forgotten effect of crossdip is that -depending on the cross-offset distribution -it can not only de-focus and smear reflections, but also shift them in time. Most existing crossdip routines rely on a slant 30 stack approach for the correction which has the drawback that shifted reflections will appear twice in the stack. Within our In the second part of this paper, we presented results of reprocessing data from the Burträsk profile using our new module.