Barents Sea carbonate reservoir detailed
Norsk Hydro ASA
Today's geologists involved with hydro-carbon exploration and reservoir characterization have to face an increasing number of seismic volumes. The inter-preter is expected to comprehend data sets processed with different parameters or multi-sets of time-lapse seismic and a plethora of attributes.
These seismic volumes reveal different features, and synthesizing all available information is challenging. Three-dimensional (3D) classification of seismic facies is crucial for seismic interpretation of sedimentary bodies and detailed reservoir characterization. The seismic facies are mapped using 3D texture attributes.
We present a new procedure for combining this information into a 3D geological model using 3D seismic facies classification with a neural network supervised classification algorithm. The model describes 3D seismic facies in opposition to the way seismic facies analysis is done presently (i.e. by classification of the waveform bounded between stratal surfaces). The advantages of this approach are:
- It provides detailed 3D seismic facies classification compared to 2D classification results (cross-sections or maps)
- The automation provides fast turnaround and reproducible results
- The 3D mapping of sedimentary bodies reveals new geological information compared to the original model built on well data and outcrop analogs.
Seismic facies analysis
The reflection of seismic waves within sedimentary rock bodies produces an image of their external shape and their internal configuration or texture. The analysis of these shapes and textures is referred as to seismic facies.
Seismic facies examples used as training data or calibration images.
The parameters and signal characteristics of a group of reflectors typify a seismic facies, which differs from a neighboring set of reflectors. The lateral and vertical distribution and associations of the seismic facies reveal patterns that need to be interpreted.
3D texture attributes
Automated 3D seismic facies mapping must honor the stratigraphic and structural framework of the seismic data. Similarly to the interpreter's eye that follows the local orientation of reflectors and stratigraphy across a seismic section, the texture attributes have to be dip-steered.
The dip-steering constraint is achieved by using 3D attributes that follow and capture the seismic patterns. These attributes are referred to as 3D texture attributes, since they are able to describe the reflector-geometry in a small 3D neighborhood (local orientation-guided multitrace attribute). The combination of the 3D texture attributes, containing orientation and/or continuity information, identifies the stratigraphic patterns from the seismic data.
A) Cross-section of the seismic data. B) Corresponding cross-section in the classification cube.
Texture attributes are subdivided into two groups. The first includes geometric attributes that capture the reflector orientation or the reflector continuity information. The second defines signal attributes capturing amplitude and frequency information. For geometric attributes, our dip and azimuth estimation (local orientation estimation) approach is based on three steps:
- The gradient vector is estimated and calculated
- The local gradient covariance matrix is calculated
- A principal component analysis is done giving the dominating orientation normal to the local reflection dip and azimuth.
The variance attribute uses the local variance as a measure of signal unconformity or discontinuity. The variance is computed for each voxel in small horizontal sub-slices. If the slice is within an unbroken reflection layer, the amplitude variance is small. Amplitude changes due to discontinuous horizon, however, result in a large variance.
The edge enhancement attribute enhances spatial discontinuities by measuring changes in the signal amplitude. The edge enhancement attribute uses the local dip estimate of the reflection layers. The local dip estimate represents a plane, and by projecting the vector with derivatives, onto this plane, changes that are nearly perpendicular to the reflector produce vectors with small magnitude. Changes in the direction of the reflector will produce vectors with larger magnitudes.
The flatness attribute is derived from the local dip and azimuth estimates. It indicates how flat or planar the reflectors are by measuring the local variance in the orientation field.
Signal attributes are generated from the original seismic cube. The volume reflection spectrum (VRS) attributes realize spectral analysis of the reflectivity response for each seismic trace. Each trace is characterized in terms of its eigenvalue (spectral attribute) and the associated eigenvector (orthogonal polynomial) to approximate the reflection amplitude along the trace in a least square sense. For texture mapping, a set of discrete spectral VRS coefficients is combined into a composite spectral representation.
The gradient-based attributes are generic 3D-texture attributes, where the gradient for each voxel is calculated by the spatial derivatives in t/d, x, and y directions from the spectral decomposition.
Neural network classification
Automated 3D mapping of seismic textures requires a classification step, where geometric and signal attributes information are combined into a 3D seismic facies cube. The supervised classification process allows selecting calibration images on different seismic facies within the volume of interest and using these data to train the artificial neural network.
Neural networks have been often used to analyze data and to recognize patterns within data. In supervised classification, both the inputs and the outputs are provided. The network then processes the inputs and compares its resulting outputs against the desired outputs. Errors are then propagated back through the system to adjust and update the weights, which control the network. This process occurs over and over as the weights are continually tweaked.
Training data, or calibration images, enable the training of the neural network. The results of classification come together with probability cubes or confidence measures for each class.
The interpreter picks the training data by digitizing different examples of seismic facies, which serve as calibration images for the neural network. Guided by the confidence measures, the interpreter can refine the training step by deletion and/or addition of calibration images until a satisfactory result is reached, i.e., similar to what would be obtained by doing a manual seismic facies analysis on a 2D cross-section). When the neural network has been properly trained, the classification can be run on the full volume of interest.
The workflow involves several steps. 3D texture attributes and signal attributes that capture both seismic textures or patterns and amplitude information are generated. Next, training data on the seismic facies is used. Then, both training data and attribute sets are classified and optimized. Finally, the 3D classification is run on the volume of interest.
The interpretation phase involves calibration with well data because of the non-unique relationship between seismic facies, lithofacies, and rock properties. By verifying if the facies associations and map patterns have geological and sedimentological meaning, the interpreter works at a higher level.
Finding carbonate buildups
Reefs or buildups represent significant potential hydrocarbon reservoirs in carbonate deposits. 3D geometry and distribution of reefs and buildups on carbonate platforms or ramps are controlled by many different variables including accommodation space (eustacy, subsidence, and local tectonics), sedimentation rate, water energy, nutrient availability, and light.
Understanding, modeling, and predicting carbonate geometries are challenging tasks. Most outcrop analogs do not permit 3D inspection and modeling of the architecture.
In the Norwegian Barents Sea, the Late Paleozoic deposits consist mainly of shallow marine, locally evaporitic, dolomitic carbonate sediments with seismic scale carbonate buildups. In this case study, automated seismic facies mapping performed an objective and fast 3D identification of the carbonate buildups that would have been very challenging if done manually. The method allowed the interpreter to quickly isolate a volume of interest where the buildups are concealed and to focus the interpretation work directly on the potential hydrocarbon reservoirs.
In this example, the training data were calibrated on six seismic facies, including parallel high amplitude (PHA), parallel low amplitude (PLA), subparallel reflectors (SR), dipping discontinuous reflectors (DDR), chaotic reflectors (CAR) and wavy reflectors (WR).
Classification results and interpretation
The seismic facies classification using interpreter-trained neural networks allowed the interpreter to perform 3D mapping of the carbonate buildups, which were checked against the original seismic data. A cross-section in the classification cube reveals the internal architecture of the buildups including the dominant vertical stacking (aggradation) of the buildups, the lateral connections in the up-dip direction (landward stepping), and the connections between the main buildups.
The classification cube allows the interpreter to analyze the facies distribution in any cross-sections, but particularly along/ above/below geological boundaries, stratal surfaces or horizons. Stratal map views and 3D views also show that the Paleozoic carbonate buildups of the Barents Sea are not isolated mounds but are connected into a complex mosaic of built-up ridges and semi-enclosed polygonal basins/lagoons.
Continuous linear ridges of the buildups running along the ramp are interrupted in a strike-slip manner by main faults running across the ramp. Detailed analysis of the seismic facies classification cube and the original seismic data confirmed the correlation between the location of pre- and syn-sedimentary faults, and the location of carbonate buildups on the ramp.
In addition, stratal map views generated along and below the top of the Paleozoic sequence provided a significant contribution to the identification and interpretation of karstified buildups, sinkholes, and paleo-caverns that developed along and below depositional sequence boundaries. The distribution of these features strongly influences well planning decisions. In a later stage, these geobodies can be calibrated and populated with rock properties using well data or other types of attributes (AVO, impedance cube etc.) and forwarded to 3D property models.
Seismic facies mapping applied to the Barents Sea case study revealed unexpected complexity of the carbonate buildup architecture and a significant contribution to the identification of karst and sinkholes. This approach offers new geological information for predicting and modeling reef distribution where 3D model constraints are poor or unavailable.
In the case of discrete objects, such as carbonate buildups or channels where lateral distribution is hard to predict and model even with stochastic approaches, automatic seismic facies mapping provides detailed 3D geometric information. The mapping of carbonate buildups from seismic data is a step forward in texture recognition for hydrocarbon exploration and reservoir characterization.
Elvebakk, G., Hunt, D., and Stemmerik, L., 2002, "From isolated buildups to buildup mosaics: 3D seismic sheds new light on upper Carboniferous-Permian fault controlled carbonate buildups, Norwegian Barents Sea," Sedimentary Geology, 152, 7-17.
Randen, T.; Monsen, E.; Abrahamsen, A.; Hansen, J. O.; Schlaf, J., and Sonneland, L., 2000, "Three-Dimensional Texture Attributes for Seismic Data Analysis," Ann. Int. Mtg., Soc. Expl. Geophys., Exp. Abstr.
Sonneland, L., et al, 1994a, Volume Reflection Spectral Analysis. Schlumberger Geco-Prakla, internal report.
Sonneland, L., et al., 1994b, 3D Model-based Bayesian Classification, EAGE.
We acknowledge the support from the EU, IST-1999-20500 and IST-1999-29034 research programs, and wish to thank Norsk-Hydro, Statoil, AGIP, and Fortum for permission to publish data.
For information, contact: Alexis Carrillat, Schlumberger Stavanger Research, telephone: (+47) 51 94 65 13 / 51 94 60 01; email: firstname.lastname@example.org.