A comparison of popular neural network facies-classification schemes

Abstract

Two learning networks are commonly used for seismic facies classification: unsupervised and supervised neural networks. While unsupervised neural networks have proven to be effective processes for macro- and mesoscale depositional facies characterizations, published results from supervised neural networks have more often demonstrated effective reservoir-scale characterization studies by using neural network mapping rather than neural network classification methods. Neural networks are sophisticated techniques that reduce data dimensionality and assist in seismic interpretation for exploration, exploitation, and production projects. In essence, neural networks for seismic interpretation map multiple seismic attributes to an output attribute or attributes that incorporate combined facets of the input data sets that yield a more accurate subsurface interpretation. Selecting an appropriate neural network is crucial to maximize interpretation benefits, yielding an accurate representation of the subsurface such that it can be used meaningfully to effectively support exploration and production efforts. These two classification techniques are compared using Oligocene Catahoula oil-bearing sands overlying a shallow salt structure along the Gulf Coast, where well control is abundant and used to ground-truth the seismic classifications of shale, brine sands, and oil-sand extent and thickness. Extraction of the seismic facies traces at various wells within the 3D volume allows direct comparison of the two techniques with known geology, permitting an unbiased evaluation and demonstrating the advantages of supervised neural networks for the classification of pertinent geologies. The advantages of the supervised network are demonstrated further by horizon and stratal slices that properly identify productive fault blocks while minimizing false positives of single seismic attributes.