Neural network classification method helps seismic interval interpretation

Sunit K. Addy
CGG Americas Inc.
Houston

The shape of a seismic wiggle trace and its geologic use can help classify a seismic interval parallel to a horizon in a 3D data volume.

A new technology that uses neural network methods for pattern recognition successfully applied this classification method to geological interpretation of seismic data in central South Texas.

Trace shape is a fundamental property of seismic data, since traces contain all relevant information-such as reflection patterns, phase, frequency, and amplitude. A map showing the distribution of similar trace shapes is like a facies map showing similar geologic features. Such a map is described here as a "seismic facies map."

Software developed by CGG-Petrosystems generates the maps. It uses a neural network technology, patented by and licensed from Elf Aquitaine, to create a set of model traces by an unsupervised learning process on a subset of data, then organizes the model traces showing progressive changes in them.

CGG used this method to produce facies maps from the 3D survey in central South Texas. Among results of the project:

• In an Upper Wilcox seismic facies map we were able to identify many features of a fluvial deltaic river system, such as fluvial channel fills, braided deposits, point bars, crevasse splays, mouth bars, fan delta, and abandoned delta.
• The seismic facies map helped determine the extent of Lower Wilcox onlapping horizons on the base of the Lavaca channel in a finding confirmed by onlap mapping.
• Using facies and other attribute maps, we were able to identify broad areas of porosity development in Cretaceous Edwards limestone, which is a major hydrocarbon producer in this area.
• Facies maps of an interval around Sligo limestone showed a linear northeast-southwest trending reef at the shelf edge for the entire 14 km length of the survey. We were also able to interpret parallel interior reefs, back reef sediment-filled basins, patchy reefs, and tidal flats from the facies maps.

The process is extremely fast, and many facies classifications can be performed in a relatively short time for several horizon intervals. This enables the interpreter to effectively scan through the data quickly and to concentrate on areas of significance, where the classification process can be repeated in prospect-size scale.

Pattern recognition

The concept of tying seismic facies with well information and predicting lithology away from the wells using seismic facies maps has been used successfully in several instances. ¹ But in most cases seismic facies maps come from visually recognizing reflection patterns in seismic data and classifying them into descriptive classes, such as continuous, sigmoid, discontinuous, hammocky, and transparent.

This process is painstakingly slow and interpretative. The maps are manually drawn and are subject to change from interpreter to interpreter.

The pattern-recognition capability of neural net overcomes the problem. A seismic facies map can be made into a geologic facies map if well information is available or by predicting the geologic facies from sequence stratigraphic and other geologic considerations.

The importance of using seismic data as a start for a geologic facies map is that such a map gives the overall variability of the facies. This may not be possible if only well data are used.

Methodology

The 3D survey for the project reported here was shot by CGG in 1995 around Hallettsville in central Lavaca County, Texas. The survey extends over 114 sq km in an area with several oil and gas fields.

Stratimagic, the latest CGG-Petrosystems 3D seismic stratigraphic interpretation software, interprets a horizon of interest through a trace peak, trough, or zero-crossing using an advanced model-based 3D propagator and edits the interpretation using a horizon auto tracker. No smoothing function is applied to the auto-tracked horizon.

A constant interval parallel to the horizon is created containing the features of interest, which should appear on the facies map. The thickness of this interval could vary from a few to several hundred milliseconds; normally, intervals of 40-150 ms yield interpretable facies maps. For small prospect areas intervals of less than 40 ms may be appropriate.

The objective is to generate a facies map of this interval based upon shapes of the seismic traces. There are two steps in the process.

First, the various shapes of the traces within the interval are analyzed by neural network, which generates a series of synthetic traces that best represents the diversity of shapes in the interval and organizes them in a progressive manner assigning each synthetic trace a color and a number.

The number of classes is determined by the user.²

Second, once the synthetic traces are available, each trace in the interval is matched against all the synthetic traces and is assigned the number and color of the synthetic trace to which it has the highest degree of correlation.

The resulting seismic facies map is a map of similarity between actual traces and a number of synthetic traces. This is an unpiloted method and requires no acoustic modeling or well input. However, if needed, traces from synthetic seismograms from wells can be used as model traces. The seismic facies along the traverse of a seismic line can be projected on the line.

Attribute facies maps can be generated from various attribute measurements, such as many forms of amplitude attributes and frequency attributes with the same neural net technology.

Such maps can be created not only for constant interval but also for variable interval.

Upper Wilcox example

Figs. 1-9 show the technique at work mapping an Upper Wilcox channel system from the 3D survey in South Texas.

In Fig. 1 [129,155 bytes], an inline and crossline through the survey show a channel system in which the top channel and base channel horizons have been interpreted.

Fig. 2a [119,479 bytes] is a time structure map of the channel top horizon, the reference horizon. The channel base horizon is 10-90 ms below this event (Fig. 2b). Two small, meandering rivers can be seen joining to form a small radiating delta.

The log portion showing more than 200 ft of channel fill in Fig. 3 [52,229 bytes] is from Mobil Spanihel 1. A constant interval of 100 ms, starting 10 ms below the channel top horizon, is created, as shown on Fig. 1b. This interval parallel to the channel top horizon contains the channel system and was analyzed by neural net for classification of the traces according to their shapes.

In the first step, a set of 12 model traces belonging to 12 classes was created from a subset of the data (Fig. 4 [81,092 bytes]). The next step was to correlate each of the actual traces within the 100 ms interval with all the model traces. The result is a seismic facies map (Fig. 5 [107,445 bytes]).

The correlation map in Fig. 6 [101,992 bytes] shows that, in general, within the channels the correlation between the model traces and the actual traces is poor, suggesting that the channel area has more diversity than the four or five facies classes shown here. The facies map can also be projected on the seismic line to show the geologic significance of a particular facies type (Fig. 7 [86,426 bytes]).

The interpretation of the facies map shows many interesting physiographic features, such as point bars, braided river bed, levee, crevasse splays, flats, fan delta, and abandoned delta. These appear on the mixed map in Fig. 8 [143,592 bytes], generated by superimposing the average amplitude map (in a black and white scale with white as high amplitude) on the seismic facies map to enhance some of the subtle features. Such a map generated in the early part of the exploration program helps the interpreter to quickly focus on important features.

Since a single attribute map may not adequately display all details, it is a common practice to create many such attribute maps and examine them individually. This task can be time-consuming and often leads to confusion as to which of the attribute maps provides the most and most-correct information.

Attribute maps can also be combined through use of the neural net. Fig. 9 [97,753 bytes] is an example. Here, a variable interval between the channel top and bottom were used to generate several amplitude attributes, such as average sample amplitude, their standard deviation, third moment, fourth moment, maximum peak, maximum trough, and frequency.

Measured attribute values and isochron data were used for neural net classification. The map shows the distribution of high amplitude events in much greater detail by the red facies, identifying several smaller point bars, longitudinal mouth bars in the delta, and other flats in the southern channel.

Well information suggests this facies indicates possibly high sand content. Normally, an attribute facies map is studied along with a seismic facies map to obtain the most information.

Lower Wilcox example

The Lavaca channel was developed during a low stand of the sea level in Eocene time and was further modified by submarine processes during subsequent drowning to become a fairly broad marine channel.

The base of the Lavaca channel is a profound unconformity as shown in Fig. 10a [223,831 bytes], an inline from the southern part of the area. In the study area this horizon forms a high dipping to the northeast and southeast (Fig. 11 [51,673 bytes]). A shelf edge canyon fill, also known as Hallettsville complex, was deposited in the Lavaca channel during the high stand. Within this mud-dominated canyon fill complex Lower Wilcox sands are found to onlap on the base channel and are good exploration targets.

Fig. 12 [198,321 bytes] is a portion of a Mobil Spanihel 1 log showing sand (marked by an arrow) at 9,820-70 ft at the base of the fill, which tested 1,030 Mcfd of gas and oil.³ This sand corresponds to the upper onlap in Fig. 10a. We observed similar onlaps downdip as shown by onlap arrows on the time structure map (Fig. 11).

Our objective is to map such onlapping horizons with a seismic facies map. An interval of 110 ms was created above the Base Lavaca channel to include the onlapping horizons (Fig. 10). A neural net classification of the interval involving 10 classes yielded the seismic facies map shown in Fig. 13 [89,329 bytes].

Using well information and a few limited onlap terminations, we interpreted the overall pattern observed in the facies map. Then the interpretation was confirmed by additional onlap interpretation. The series of concentric facies around the high reflects the onlap of individual horizons of the Lower Wilcox Hallettsville complex on the slope of the Lavaca channel. Plotting the seismic facies on the line makes it easy to correlate onlaps with facies types (Fig. 10b).

Edwards Limestone example

Cretaceous Edwards limestone in central Lavaca County shows porosity development in the upper 350 ft, believed to result from karsting and fracturing. Prediction of porosity and its lateral continuity is a major exploration issue since several fields, such as Word and Hallettsville, produce gas from Upper Edwards.

Fig. 14a [165,930 bytes] is an inline showing the Edwards reflector and an interval of 80 ms created for neural net classification with 10 classes. The facies map (Fig. 16 [176,251 bytes]) shows a swath of different facies (red, yellow, and others) through the center of the survey running northeast-southwest along the shelf edge (F1 in Fig. 15 [102,540 bytes]) extending several kilometers on both sides.

Comparison of the facies types and average porosities in a few wells reported by Weathers⁴ reveals a general trend in which wells within the red and yellow swath show higher porosity. Best porosity is observed when the well is within the red-yellow swath and next to a fault, as in WW12.

Fig. 17a [146,926 bytes] is a mixed map of the Edwards seismic facies map and the Edwards horizon dip map. The dip map particularly shows the possible extent of karsting on either side of the shelf edge, as indicated by sinuous boundaries within which the red and yellow facies are contained.

The author has suggested that the porosity was a result of karsting at the shelf edge during a low stand of the Edwards sea (Fig. 17b).⁵

In Fig. 14b, the seismic facies and the average amplitudes plotted above and below the interval, the subtle difference in the Edwards reflector can be observed between areas of no karst and karst development.

Sligo reef example

The outer edge of the Lower Cretaceous carbonate platform around the Gulf of Mexico shows reef development at the shelf edge. This trend has been named Stuart City, Edwards, Glen Rose, and Sligo reef trend, depending upon location.

In the study area, Sligo limestone is overlain by a Pearsall shale about 250 ft thick, which in turn is overlain by about 2,000 ft of Edwards limestone and 300 ft of Austin chalk, successively. These horizons are shown in Fig. 22 [188,654 bytes], an inline (A-A') from the southern part of the survey area (Fig. 18 [55,717 bytes]). Most Lower Cretaceous production comes from the Edwards formation (about 13,200 ft); and several wells have been drilled to Edwards. Wells penetrating Sligo (about 15,000 ft), on the other hand, are few.

Hydrocarbon production from this reef trend has been documented in several fields, such as Fairway field and Fort Trinidad field in East Texas,^{6 7} but in central Lavaca County Sligo reef has not been tested properly. Here, the Sligo reef play is considered as a frontier type play. Details of the reef geometry will be reported here based on the seismic facies map.

Fig. 18 is a time structure map of Sligo. An interval of 90 ms below the Pearsall horizon was created (also shown on Fig. 23). This interval is about 500 ft thick and includes Pearsall shale and the upper part of Sligo.

Various volume attribute maps, such as average sample amplitude, maximum peak, and maximum trough, were also created. A seismic facies map of this 90 ms interval was generated by the neural network process. The computed model traces are shown in Fig. 19 [89,738 bytes].

Fig. 20 [147,352 bytes] is the seismic facies map of the defined interval mixed with the azimuth (of the dip) map of the Sligo horizon. The interpretation of this map involved constant referral to the seismic sections and attention to reef geometry for a high stand situation. Combined with the seismic sections and patterns in the facies map, our interpretation shows the shelf edge, fringing reef, inner reef, patchy reef, back reef basins, lagoons, tidal flats, upper slope with or without reef debris, and other features.

This map shows various subtle features and their lateral extent, which are otherwise not visible in a structure map (Fig. 18). It is clear that a less distinct inner reef parallels the fringing reef on the shelf. Between these two reef trends there is a back reef basin, which shows thicker accumulation of sediments in the south. West of the inner reef there is another back reef basin showing similar, thicker sediments to the south.

We have also interpreted patchy reefs, intervening basins, and tidal flats on the inner shelf.

Generation of porosity in a reef is a key exploration question. The Sligo reef trend has been classified by Read as an accretionary rimmed shelf margin with a gentle slope.⁸ In this type of a reef, normally the reef growth keeps a balance with the sea level rise, causing upward growth of the reef. However, the growth is hindered when the reef is exposed; in such cases, secondary porosity is developed due to weathering.

Amplitude maps often indicate such porosity zones. An average interval amplitude map (Fig. 21 [86,138 bytes]) shows lowering of amplitude at several places on the reefs identified by the seismic facies map. Like the seismic facies stripe shown above the interval in the seismic line, the amplitude can also be shown as a stripe below the interval for visual correlation (Fig. 22).

In a conventional interpretation work flow using structure and isochron maps, the interpreter would have to guess about many of the features shown in Fig. 20 . With the help of the seismic facies map the interpretation is much easier.

Quick tool

The neural net-generated seismic facies map is a quick exploration tool. Such maps can answer key questions at the beginning of the exploration phase.

Within a very short time in a 3D survey, we were able to interpret many sedimentary features of a fluvial deltaic river system in Upper Wilcox, map the extent of Lower Wilcox onlapping horizons, identify broad areas of porosity development in Cretaceous Edwards limestone (which is a major hydrocarbon producer in this area), and interpret the Sligo reef trend and associated fringing reef, back reef basins, inner reefs, and other features.

Acknowledgment

The author thanks Graham Cain, CGG-Petrosystems, USA, for the use of the software, Jim Tucker of Integrated Studies for encouragement and support, and CGG for the use of the data and permission to publish it. Special thanks to Philip Neri, CGG, France for introducing Stratimagic to me.

References

1. Addy, S.K., Behrens, E.W., Haines, T.R., Shirley, D.J., Worzel, J.L., "High frequency subbottom reflection types and lithologic and physical properties of sediments," Marine Geotechnology, Vol. 5, 1982, pp. 27-49.
2. Kohonen, T., "Self organization and associative memory," Springer-Verlag, 1984.
3. Devine, P.E., Wheeler, D.M., "Correlation, interpretation, and exploration potential of Lower Wilcox valley fill sequences, Colorado and Lavaca Counties, Texas," Trans. Gulf Coast Asso. Of Geol. Soc., Vol. 39, 1989, pp. 57-74.
4. Weathers, R.L., "Reservoir description using acoustic impedance-Hallettsville 3D," SEG Tech. Prog., Vol. 66, 1996, p. 762.
5. Addy, S.K., "Attribute analysis in Edwards limestone in Lavaca County, Texas," SEG Tech. Prog., Vol. 67, 1997.
6. Terriere, R.T., "Geology of Fairway field, East Texas," in "North American Oil & Gas Fields," AAPG Mem. 24, pp. 157-176.
7. Pampe, C.F., "The Fort Trinidad Glen Rose Field, Contr. to the Geology of South Texas," So. Tx. Geol. Soc., 1967, pp. 25-30.
8. Read, J.F., "Carbonate platform facies model," AAPG Bull., Vol. 59, 1985, pp. 1-21.

The Author