The application of multivariate statistics and neural networks to the prediction of reservoir parameters using seismic attributes

Brian H. Russell

In this dissertation, I develop a number of new ideas for the statistical determination of reservoir parameters using seismic attributes. These ideas combine the classical techniques of multivariate statistics and the more recent methods of neural network analysis. I apply these techniques to both full seismic volumes and to maps derived from intervals averaged through these volumes, largely using the Blackfoot dataset from central Alberta. I show that multilinear regression often provides too simple a solution to the parameter estimation problem, and that the traditional feedforward neural network often provides a solution that is overly complex. My proposed solution is to use radial basis function neural networks for the prediction of reservoir parameters, since this approach combines the power of the multilinear regression technique with the nonlinearity of neural networks. I also show how the radial basis function neural network can be considered as a generalization of the generalized regression neural network, which has been previously used in this type of parameter estimation. My conclusions are illustrated using both an AVO classification problem and the Blackfoot seismic and well log dataset.

In addition to the application of the radial basis function neural network to the prediction of reservoir parameters, several new ideas are presented for the analysis of well log and seismic data. First, I derive an improved regression formula for the prediction of S-wave sonic logs from combinations of other logs. Second, I apply a new approach to data clustering, which I call Mahalanobis clustering, to the interpretation of AVO crossplots and to the delineation of optimal clusters for the radial basis function neural network with centres method. Finally, I develop a new approach to map analysis that combines geostatistics with multiattribute transforms. collocated cokriging technique. This technique uses multivariate statistics and neural networks to improve the secondary dataset used in the collocated cokriging technique.