Multivariate statistical analysis of seismic data applied to the delineation of reservoir facies in the Birkhead Formation, Sturt Field, Lake Hope Block

Abstract

The delineation of seismically thin reservoirs is a common problem in the exploration for and development of hydrocarbon reserves in many basins. Reservoirs which are thinner than a quarter of the dominant wavelength of the incident wavelet cannot be resolved using time difference measurements made on the seismic waveform. The "thin bed" can only be inferred by waveform analysis. Most "thin bed" studies involve the examination of only a few attributes, for example amplitude and./or frequency of the seismic data. This thesis however, applies multivariate statistics, which can deal quantitatively with as many attributes as the user cares to define. The target chosen for this study was the thin reservoir sandstones within the Birkhead Formation of the Sturt Field in the Eromanga Basin, South Australia. These sands form combination structural-stratigraphic traps which are very difficult exploration targets. Before the statistical analysis, l-D and 1.5-D modelling was performed and this demonstrated that some of the thirty-four parameters measured from the Birkhead seismic signature should show measurable variations between the reservoir and non-reservoir facies. The parameters were measured from the trace amplitude, autocorrelation, power spectrum, complex amplitude and instantaneous frequency. Modelling also provided an understanding of the way in which geologic variations affected the seismic attributes from which the parameters were calculated. The next step was applied to both the model and field data. It was modified from the procedure reported by Dumay and Fournier (1988) and involved selecting a set of reference traces around wells where the geology was known. Each set of traces represented either the reservoir facies or a non-reservoir facies. Principle Component Analysis (PCA) was then applied to see if the seismic attribute differences could be detected statistically. Many of the thirty-four variables showed some interdependence and were therefore omitted. A total of six significant variables remained in the analysis. For the model data there were two parameters from the autocorrelation, one from the power spectrum and two from the complex amplitude. For the real data these numbers were three, one and two respectively. Discrimination between the reservoir and non-reservoir environments proved possible for the model and real data. The discriminant functions defined in the previous step were used to classify traces away from the well control where the geology, (in the real data) is unknown. Classification of the model data showed that reservoir sands could be predicted down to a minimum thickness of l5ft. Classification of the real data provided a map of the predicted reservoir facies distribution for the whole study area. At the conclusion of this study, it is not yet possible to assert that the method is reliable due to a lack of drilling in critical locations to test its predictions. However, modelling provides strong evidence that the channel sand facies is statistically distinct from the non-reservoir facies. The similarity between seismic attribute patterns for the real data and model data adds confidence in the use of this multivariate statistical method as an interpretative tool.