GNGTS 2016 - Atti del 35° Convegno Nazionale

GNGTS 2016 S essione 3.1 489 section that intersects the well that have been previously used to compute the synthetic data. This section represents for each CMP position the 1D vertical facies profile determined from the MAP solution of AVA inversion. Note the overestimation of brine sand intervals and gas sand intervals associated with the DA and the NN approaches. The comparison between the results produced by these two classification algorithms and the facies estimated by two Bayesian classification methods makes even clearer the importance of inserting a-priori information in the analyzed case. In particular, note that the MC approach returns more plausible vertical succession of litho-fluid facies and less scattered results than to the standard Bayesian method. However, all the four methods have been able to correctly identify the main gas sand interval that is located between cross-line 8420 and 8440 at 2.46 s. Similarly to the synthetic example we use the well information and the classification results obtained in correspondence of the well location to derive the confusion matrix for each classification method. Fig. 3b shows that DA method overestimates the occurrence of brine sand layers and underestimates the gas saturated intervals. More in particular, this approach tends to interpret the 40% of actual shale intervals as brine sand layers, whereas the 25% of gas intervals are misclassified as brine sands. The NN method tends to interpret the 50% of brine saturated sands as gas saturated, whereas the 20% of shales are interpreted as brine sands. The misclassification errors are lower for the two Bayesian methods and particularly for the MC approach that returns a confusion matrix with very low off-diagonal terms. Conclusions. We tested four different approaches to classify the P-wave and S-wave impedances estimated from pre-stack seismic data by means of a Bayesian AVA inversion. The significant depth at which the reservoir zone is located produces a significant overlap between the Ip and Is values of brine sands and gas sands that makes the discrimination between these two litho-fluid classes particularly challenging. Moreover, the field data classification is made even more problematic by the very low resolution of the seismic data. In this not favorable context the consideration of any kind of a-priori information in the classification process revealed to be the crucial aspect to derive a reliable facies classification. In particular, just the standard Bayesian approach that simply takes into account the a-priori information about the overall proportions of facies around the investigated interval, produces final results in which the actual proportion of facies is closely respected. If a 1D Markov chain prior model is additionally incorporated into the Bayesian classification procedure, the predicted facies profiles also show a physically plausible vertical succession of litho-fluid classes. Acknowledgments . The authors wish to thank EDISON for the permission to publish this work. References Avseth P., Mukerji T. and Mavko G. 2005: Quantitative seismic interpretation: Applying rock physics tools to reduce interpretation risk . Cambridge University Press. Buland A. and Omre H., 2003: Bayesian linearized AVO inversion . Geophysics, 68(1), 185-198. Larsen A.L., Ulvmoen M., Omre H. and Buland A. 2006: Bayesian lithology/fluid prediction and simulation on the basis of a Markov-chain prior model. Geophysics, 71(5), R69-R78.