Some considerations for quantifying sensitivity using 4D seismic data

Whilst the seismic–based approach appears promising for quantitative measures of the reservoir’s sensitivity, there are challenges with 4D seismic data that should be of note, both for amplitudes and time-shifts.

4D seismic amplitudes

Seismic amplitudes are affected by acquisition geometry/shot variation, equipment, propagation effects and processing algorithms/workflows, and, as such, are non-unique (Sheriff, 1975; Simm and Bacon, 2014) (see also Figures 1-3 and 1-4, in Section 1.1). For the interpretability of 4D seismic data, one criterion is data repeatability (Kragh and Christie 2002) which is believed to be affected mainly by acquisition-geometry differences, near-surface conditions, environmental conditions, noise, geology etc. as discussed in Johnston, 2013. Whilst co-processing algorithms and parameters for 4D processing (Helgerud et al., 2011) aim to preserve the 4D signal (amplitude and phase) and improve repeatability, the 4D signal is not always preserved in processing. Migration, in particular, has a complex averaging effect on the 4D signature and may

also introduce additional noise (Canning 2010) that will impact quantitative 4D amplitude studies. More important is the use of appropriate velocity models for 4D imaging. Brain et al. (2009) reveals better 4D signal preservation and interpretation using a pre-stack depth migrated data (which requires more accurate velocity models) compared to that of a time migrated data. In addition, intra-survey reservoir fluctuations results in a lack of accuracy in the measured amplitude response (in post-stack 4D seismic data) to reservoir changes as they are found to be in error (Omofoma and

MacBeth, 2015; see also Chapters 4 and 5). Such irreconcilable spatio-temporal discrepancies between the acquisition itself and the dynamic reservoir during shooting of the seismic surveys also makes small pressure or saturation induced signals difficult to detect in stacked data and may appear as noise, besides the usual factors affecting data non-repeatability. The analysis in this chapter is thus limited to areas of genuine production-induced signals caused by fluid saturation or pore pressure changes that are sufficiently large.

As is common with map-based techniques, this brings problems associated with interpreting the horizon, defining a target window, selecting a method for averaging etc. and each carries its own uncertainty.

 The root-mean-square, RMS, amplitude averaging method has been chosen for the map-based calibration. This is because the resulting RMS values are always positive irrespective of whether the reservoir reflectivity interface is represented by a trough or a peak. This will thus be appropriate for any field. The RMS amplitude is also a recognised standard for 4D seismic interpretation (Stammeijer and Hatchell, 2014). As the RMS uses windowed measurements, it is also robust with respect to noise and horizon mispicks, and is suitable for a wide range of field seismic datasets with different repeatability measures. Other methods for averaging such as Amplitude Envelop, Sum of Negative Amplitudes, and Sum of Positive Amplitudes etc. can also be used, but may not be as robust as the RMS. These other averaging methods must also be selected appropriately for a negative or positive or near-zero reflectivity interfaces.

 Appropriate time windows are required to compute the RMS amplitude maps along the top reservoir interface and this is done for both baseline and for each

monitor data individually. In this work the top reservoir picks are used because these are generally well defined and easier to pick than the bottom of the reservoir. This is most suited to seismically thin reservoirs (usually < 50 m thick) of half a wavelength. The window size for averaging the amplitude signals is straightforward as this is typically the same as half the wavelength (i.e. peak to trough). Thus, there is some tolerance allowed in the mispositioning of the window to produce a reliable RMS amplitude signal. In thick reservoirs (i.e. reservoirs represented by one or more wavelength cycles), considerations have to be given to the particular effect that can best be captured at top reservoir or at the bottom of the reservoir. Falahat et al. (2013) suggest that the effects of pressure and fluid saturation changes occupy different volumes in the reservoir. Pressure effects are likely to be laterally extensive, occupying the entire thickness of the reservoir until barriers are intercepted, whereas, water or gas saturation effects may concentrate in the deeper and shallower areas, respectively, depending on gravity, pressure gradients, mobility ratios and reservoir heterogeneity.

 It is also suggested to test different search windows and the final window choice should be selected based on improved focusing of the mapped 4D signatures and good window coverage over the reservoir interface. Larger windows will be found to be poorer focus and resolution due to too many destructive and constructive events being averaged out, but smaller windows will not capture adequately all the 4D changes. The choice of window for averaging will thus affect the magnitude of the 4D amplitude signals in the resulting maps, which adds to the non-uniqueness in amplitude sensitivity estimates.

As this study quantifies sensitivity across monitor times, it is best to use the same window size for the monitor and baseline seismic datasets belonging to a field. The RMS 4D amplitude maps are the difference between each monitor RMS map and the baseline RMS map. Differencing in the maps domain, rather than directly from volumes (from which 4D maps can then be computed) reduces the influence of residual time- shifts which are known to corrupt the amplitude information (Hughes, 2000; Alsos et al., 2009). This is common practise when the monitor and baseline seismic datasets have not been adequately time-shift corrected (Johnston, 2013). Processing of 4D

seismic data usually aims to reduce such time-shifts as much as possible. These time- shifts refer to the time misalignment between baseline and monitor traces which can be caused by production (genuine 4D signals) and noise (which is what should be removed). Time-shifts due to noise originate from non-repeatable acquisition geometries, variations in source and receiver coupling, and near-surface variations. Such near-surface variations are caused by weathering layer changes and water column changes (including those that affect the elevation of the sea surface (tides, weather and currents) and those that result in variations in water velocity (temperature and salinity changes)). Therefore, cross-equalisation and local as well as global (and trace by trace) matching filters are commonly applied to the baseline and monitor data (for example, Rickett and Lumley (2001); Kristiansen et al. (2000); Magesan et al. (2005)). Nevertheless, such time-shifts cannot be completely removed for several reasons - acquisition geometry, processing algorithms, velocity models, and parameterisation (Johnston, 2012), hence the name, residual time-shifts. For example, residual time-shifts in the range of ±1.0ms are observed above the reservoir of the Curlew D field, located in the Central North Sea and are suspected to be caused by the differences in source and receiver positioning between the surveys, but cannot be removed during processing (Fehmers et al., 2007). In such cases of poor time alignment (as is the case for the various seismic datasets in this study), the 4D amplitude signatures on maps computed separately for baseline and monitor seismic data (using the same window size), and then differenced, are better interpretable than on 4D seismic vertical sections. If the time misalignments are large between baseline and monitor times, then it is best to repick the reservoir’s horizon, or apply a bulk shift to the horizon from the baseline seismic data or adjust the window centre while keeping the same window size, before computing maps. However, within the window interval, residual time-shifts still affect the mapped RMS 4D amplitudes.

4D seismic time-shifts

Unlike 4D amplitudes which can be directly calculated from the processed seismic data simply by taking the difference between two 3D seismic amplitude volumes, or the difference between amplitude maps, time-shifts have to be estimated from the seismic volumes. The manual way of doing this is by picking the reservoir interface on baseline and monitor data, and taking the difference between the two horizons, which results in a

time-shift map. This however is affected by the consistency and accuracy of picks, which is not easy to manage even in noise-free data. Besides, it is not a very sensitive way of capturing travel-time differences, as 4D time-shifts within the reservoir interval are small (Figure 3-1) compared to seismic wave periods. Other non-manual, more sensitive, and volumetric methods for measuring time-shifts exists and are also susceptible to noise, thus, will trade spatial and temporal resolution against noise suppression. The most common way of measuring time-shifts is by cross-correlation methods (Rickett and Lumely 2001; Hatchell et al. 2003; Hall et al., 2005; Hale, 2009; Naeini, 2013). A key parameter for such methods is the size of the cross-correlation window which is not always straightforward to choose. For example, time-shift measurements over long windows may not be sufficiently accurate, whereas, estimates over short windows can be unstable as they are more affected by noise and non- repeatability. It is thus suggested to test for an appropriate window size (see for example, Williamson et al., 2007; Pazetti et al., 2016).

Other non-cross-correlation based methods include: non-rigid matching (Nickel and Sonneland, 1999) and time shift inversion based on formal Taylor expansions (Rickett et al., 2007; Williamson et al., 2007; Zabihi Naeini and Hoeber, 2008; Grandi et al., 2009; Zabihi Naeini et al., 2009; Whitcombe et al., 2010; Lie, 2011). These inversions work by minimizing an objective function obtained from the monitor-base difference seismic cubes and are non-linear ill-posed problems with non-unique solutions that require careful regularisation. Such inversions can be very effective if the time shifts are small, although the methods usually assume that the wave propagation is vertical, that the seismic traces are nearly zero offset and that changes in the reservoir are mainly influenced by velocity and not by density (Williamson et al., 2007). However, stability is to a greater extent a problem when large enough time shifts are encountered and such methods also require balanced amplitudes (Lie, 2011). In addition, in some thick reservoirs, time shifts may be large enough to cause cycle skipping (for example in Mitchell et al., 2009, where overburden time-shifts greater than 30 ms are observed), in which case, Rickett et al. (2007) emphasize that the inversion can become more non- linear. More expensive global optimisation techniques will thus be required as steepest- descent solutions will fail. Grandi et al. (2010) uses a priori information (production data, structural interpretation or other geological information) to carefully regularize the inverse problem in an iterative Gauss-Newton optimisation scheme. This they consider

a more stable inversion with better resolution which addresses stress-sensitive reservoirs with time shifts that are induced both by 4D velocity change and compaction (see also Chu et al., 2012; Zabihi Naeini et al., 2009).

Using MATLAB codes developed within ETLP, a few of these methods (Rickett et al., 2007; Hale, 2009; Whitcombe et al., 2010) have been tested to compute 4D time-shifts between observed monitor and baseline seismic data. The tests suggest that the different methods will output estimates that are of varying degrees of resolution and magnitude, but the character of the time-shift responses related to strong production-induced signals are usually consistent across the methods. Kanu et al. (2016) evaluate the accuracy of a number of these time-shifts methods and document their strengths and weaknesses. Time-shifts estimated from seismic data are therefore non-unique as the different methods suggest, and are also influenced by amplitude/phase changes (Hoeber et al., 2008).

For this study, I use the fast cross-correlation technique detailed in Hale (2009) which uses a Gaussian window to quantify vertical time-shifts between the baseline and monitor seismic volumes. Although other methods for time-shift estimation may be more accurate, their use in determining time-shift sensitivity is beyond the scope of this thesis. The cross-correlation technique was preferred mainly because of its robustness to noise and its fast computation time. Longer windows can be used to achieve better stability in noisy data and big time-shifts will still be preserved but poorer resolution means that the magnitude of the time-shifts are of lower accuracy. However, this may not present any significant impact on the time-shift interpretation (Tigrek and Hatchell, 2006; Pazetti et al., 2016), but will influence the magnitude of the quantified time-shift sensitivity. Since the quality of the cross-correlation is determined by the window size, time-shift volumes from several Gaussian windows were first evaluated for meaningful information (particularly around known well locations), lateral consistency and vertical smoothness before suitable window parameters are chosen. For consistency, the same window sizes is used for cross-correlation of the baseline and each monitor seismic volume, and was also found to be appropriate for the various field seismic datasets in this study.

Production induced 4D time-shifts are of two sources, 1) fluid saturation changes (with accompanying fluid contact movement in the reservoir) which causes seismic velocity changes and 2) changes in the stress-strain state due to pore pressure changes which cause both seismic velocity and physical thickness changes inside and outside the reservoir. For reservoir changes associated with fluid saturations, time-shifts associated with the velocity changes affects the reservoir as well as everything below it, whereas, the accompanying fluid contact movement impacts only the reservoir experiencing the fluid saturation changes. Attempts have also been made to quantify the two assuming that the effects of pore pressure changes are negligible (Aarre, 2006), and this is beyond the scope of this thesis.

For effects due to pore pressure changes, 4D seismic time-shifts capture the combined effects of velocity and thickness changes, within the reservoir interval; the former being the main contributor in sandstones. Separating the two effects remains practically challenging (Røste et al., 2006 and 2007). In this study, I make no attempt to separate time-shifts due to velocity changes from those due to thickness changes (i.e. physical compaction and dilation). The estimated time-shifts across the four offshore clastic fields in this study are thus assumed to be caused purely by velocity changes. So that a speed-up anomaly (i.e. positive time-shifts), may be caused purely by an increase in seismic velocity in the monitor time relative to baseline, and a slow-down anomaly (i.e. negative time-shifts) may be caused purely by a decrease in seismic velocity in the monitor time relative to baseline.

When interpreting the observed 4D seismic data, the time-shifts are not only caused by production but also by noise (i.e. residual time-shifts as discussed in the “4D seismic amplitudes” section above), such that the quantified time-shifts is a contribution from both sources. One ought to be cautious how processing algorithms affect the time-shifts in preserving the 4D seismic signals (Johnston, 2013).