ABSTRACT surfaces as separate objects. In the following, the steps
of this procedure will be discussed. Interpretation of faults in seismic data is today a time consuming manual task. A new method for automatic ex- traction of fault surfaces from conditioned fault enhanc- ing attributes is presented. The backbone of this process is formed by the attribute set chosen. Attributes well suited for fault detection and enhancement will be de- fined and procedures for fault surface extraction ex- plained. A case study of automatically interpreted fault surfaces will be shown. (a) (b) (c) INTRODUCTION
Reducing time from exploration to production of an oil-
field has great economical benefits. Oil companies have suggested that for each 6 months saved, 5 % of the total costs of the development of the oilfield is saved. In the exploration phase, some of the most time consuming tasks involves the geological interpretation of seismic data. This is today done manually by interpreters, and (d) much time could be saved by automating these tasks. This paper will focus fault interpretation. Figure 1 An attribute cube (b) is generated from The mapping of the fault network is of key importance in the seismic cube (a). The attribute is then condi- reservoir characterization. The objectives for the fault in- tioned, and from the resulting cube (c), the fault terpretation is partly dependent on which phase of a res- surfaces are extracted as separate objects (d). ervoir’s life-cycle the interpreters are dealing with. The exploration geologist is mostly interested in the large faults, i.e. faults with significant offsets, for identifying possible traps and getting an impression of the geologi- FAULT ENHANCING ATTRIBUTES cal history of the area of interest. In reservoir engineer- ing, subtle faults may also be of key importance. A fault Enhancing faults means enhancing discontinuities in the may be sealing or conducting, i.e. it may prevent or allow seismic data. This is, however, not straightforward, as the flow across or along the fault network. Detailed knowl- intersections between the different reflection layers con- edge of the fault system may thus provide valuable in- stitute great amplitude changes. Hence, we need to en- formation on where to place a well so that flow in the res- hance changes along the reflection layers and not or- ervoir is optimized. thogonal to them. In order to obtain this, some attributes use a local dip estimate of the reflection layers. A detailed Figure 1 illustrates the steps in the automatic fault inter- description is given in a recent paper [2]. pretation process. An interpreter would traditionally in- spect the seismic cube and map faults where the reflec- tion layers are broken and shifted. In computer vision, Dip and azimuth estimate however, this does typically not provide a feasible input for surface extraction. The first step is to create an attrib- The normal of a point on the reflection layer can be found ute cube capturing the faults by locally having high en- by calculating the gradient in that point, ergy along the fault surfaces and low energy otherwise. The resulting attribute cube will subsequently be condi- tioned in order to enable extraction of the high energy ∂x(t1, t 2, t3 ) dt1 ∂x(t , t , t ) ∇ x(t1, t 2 , t3) = 1 2 3 , dt 2 ∂x(t , t , t ) (a) (b) (c) 1 2 3 dt3 Figure 3 (a) A smooth reflector will have one dominating direction ( λmax >> λmid ≈ λmin ). (b) A with one partial derivative for each dimension. As illus- bent reflector will have two strong directions trated in Figure 2, there may be large variations in the ( λmax ≈ λmid >> λmin ). (c) A fault with a damage gradient estimates. From these variations the dominating zone will have gradients pointing in all directions orientation is computed by using principal component ( λmax ≈ λmid ≈ λmin ). analysis. The chaos attribute will not only enhance faults but also chaotic textures within the seismic ( carbonate reefs, channels, gas chimneys, etc). It can be tuned specifically for fault detection by taking the elongated, more or less vertical nature of faults into consideration. By estimating the chaos attribute within elongated vertical windows, or dip guided windows orthogonal to the dominating orien- Figure 2 There may be large variations in the dip tation, this is obtained. estimates, and we hence let the local dip estimate be the dominating dip found by principal comp o- nent analysis.
The dominating orientation is found by aggregating the
gradients into a covariance matrix, which is then decom- posed into its corresponding eigenvectors and eigenval- ues. The eigenvectors correspond to the three principal directions of the gradients involved in the covariance ma- trix with the eigenvalues indicating their magnitude. The dominating orientation is the eigenvector with the high- est eigenvalue, and this vector is chosen as our local ori- entation estimate.
Chaos attribute
The first fault attribute (refer to [2]), follow directly from
the dominating orientation analysis. Figure 3 illustrates three situations which are distinguishable by studying the value of the sorted eigenvalues, {λmax , λmid , λmin }. In cases where the local orientation estimate involves only gradients from a smooth, unbroken reflector layer as illustrated in Figure 3(a), λmax will be much larger than λmid and λmin . If the orientation estimate is taken across a fault, we will have situations as illustrated in Figure 3 (b) and (c) where we have large λmid and/or λmin . Cal- culating the ratio between λmid , λmin , and λmax , allows (a) (b)
us to detect such discontinuities.
Figure 4 (a) Seismic cube. (b) Chaos cube gener-
ated from (a). Edge enhancement attribute
A fault appears as changes in amplitude in the reflectors.
We should thus be able to enhance faults by measuring changes in the signal amplitude which is exactly what the edge enhancement attribute does. As previously discussed, the intersections between different layers comprise sharp edges and will produce large outputs by using conventional edge detection techniques. The edge enhancement attribute reduce this problem by using the local dip estimates of the reflection layers. The local dip estimate represents a plane, and by projecting the vector with the derivatives, dx (n1, n2 , n3) Figure 6 Edge enhanced cube generated from the seismic cube shown in Figure 4 (a). dn1 ∇ x = dx (n1, n2 , n3) dn2 , Variance attribute dx (n1, n2 , n3) The last attribute we present is the variance attribute. dn3 This attribute uses the local variance as a measure of onto this plane, changes which are nearly perpendicular signal unconformity. For each voxel, the local variance is to the reflector will produce vectors with small computed from horizontal sub-slices. If this slice is within magnitudes, whereas changes in the direction of the an unbroken reflection layer, the amplitude variance will reflector will produce vectors with larger magnitudes. be small whereas amplitude changes due to a fault will result in a larger variance. Next, the variance estiamte is smoothed by a vertical window and amplitude normalized. The variance attribute may furthermore be done dip insensitive by performing the variance estimation in dip corrected slices (“flow surfaces”).For (a) (b) more details on the variance attribute, refer to [1].
Figure 5 Illustration of the concept of dip guided
edge enhancement. (a) Derivatives indicating change along the reflector produces a projection with large magnitude in the dip plane, whereas (b) derivatives indicating change orthogonal to the reflector projects as a vector with small magnitude in the dip plane.
Taking the magnitude of the projected vector as the
attribute value makes this attribute dependent on the amplitude in the seismic data. Faults in areas of low amplitude will thus have a weak signature which may be hard to detect for a human interpreter. The visual Figure 7 Variance cube generated from the seismic appearance can be corrected for by applying some cube shown in Figure 4 (a). amplitude correction, but with the appropriate subsequent steps this may prove unnecessary in an ATTRIBUTE CONDITIONING automated fault extraction setting. The objective is that the values along the fault surfaces are local maximum The fault surfaces are captured as extrema surfaces values, regardless of magnitude. This attribute does not within the attribute cubes. The final step towards introduce artifacts by smoothing and picks up very small automatic interpretation is to extract these surfaces as amplitude changes (dimming effects of sub-seismic separate objects which can then be exported as resolution faults). Hence, even very subtle faults, which interpretation data to the database. Some features in the are very hard to visually detect from the seismic data, are attribute data make this a difficult task: captured. This property often makes it the preferred 1. The faults appear more like trends than well- attribute for fault extraction. Figure 6 shows an example defined, continous surfaces. of an edge enhanced cube. 2. Remains of the strata (e.g. chaotic strata) in the orthogonal to the fault. The estimates should be attribute data also comprise high attribute smoothed in order to increase reliability. values. 3. Intersecting faults are to be extracted as separate entities. SURFACE EXTRACTION 4. The data contain a high level of noise. Having performed attribute enhancement, the fault sur- We have developed a method which conditions the faces are sharp and continuous, and can be extracted as attributes so that these features are no longer a problem. connected components. We wish to be able to distin- The method is able to transform the faults into guish between the faults. By separating the fault systems continuous surfaces and remove noise. It is also able to on dip and azimuth in the conditioning step, this is ac- separate surfaces on their orientation. Hence, it is able to complished. Figure 8 shows an example where the we filter out horizontal surfaces (remains of horizons) and have separated the two fault systems in the inline direc- also separate the faults into different fault systems. The tion. Similarly, we separate fault systems in the crossline method is robust and produces very convincing results direction. (Figure 8). The surfaces are then easily extracted as connected com- ponents from these cubes and written directly to the da- tabase as interpretation data. Figure 9 shows extracted surfaces visualized in an interpretation tool.
Figure 9 The extracted fault surfaces visualized in
a sub volume of the seismic cube.
Figure 8 Result of conditioning the edge en-
hancement cube (Figure 6). The conditioned cube SUMMARY AND CONCLUSIONS is split up into separate fault systems, here illus- In this paper, we have presented the steps in a method trated by the two dip systems for faults propa- for performing fully automatic fault interpretation. We gating in the inline direction. use fault enhancing attributes to detect faults. The at- tributes have different properties and which attribute to The surfaces resulting from the fault enhancement choose is usually data dependent. If a very detailed in- method are not always sharp. Further conditioning of the terpretation is desired, the edge enhanced attribute is result is performed by thinning. The thinning operation usually the preferred attribute. The attributes produce will for each voxel check whether the current voxel is the high energy outputs along the fault surfaces, but rela- peak value of a neighborhood of n voxels forming a line tively little connectivity within the surfaces. Intersecting perpendicular or nearly perpendicular to the fault. If so, it faults and noise make surface extraction from this data is retained, otherwise it will be discarded. As a result, the difficult. This is handled by conditioning the attribute surfaces will be thinned. In order to identify the cubes upon extracting the surfaces. The conditioning al- neighborhood in which to look for a peak value, an lows separation of the fault surfaces based on dip and estimate the normal to the fault in the voxel of interest is azimuth, such that cubes with non-intersecting faults can needed. One way to make this estimate is to use the be created. From these, the surfaces are extracted as con- orientation of the gradient vector after projection onto nected components. the orientation plane, as described in connection with the edge enhancement attribute. This vector will generally be REFERENCES [0] T. Randen, PCT Patent Application No PCT/IB99/01040 AUTOMATED STRATIGRAPHIC AND FAULT INTERPRETATION 14.0110/0151 [1] P. Van Bemmel, R. Pepper, “Seismic signal process- ing method and apparatus for generating a cube of variance values”, Patent US06151555 [2] T. Randen, E. Monsen, et al. , “Three-Dimensional Texture Attributes for Seismic Data Analysis”, Ann. Int. Mtg., Soc. Expl. Geophys., Exp. Abstr, 2000.