Seismic attributes extracted from 3-D seismic data are important for us to integrate the subsurface structure and to predict the hydrocarbon reservoir. Numerous methods have been developed for reservoir prediction based on neural network such as BP and SOM using seismic attributes. This paper proposes a seismic attributes selection method which choose the useful ones from the various kinds of seismic attributes. The selected attributes will be then used for regression to predict the hydrocarbon reservoir. A real seismic data example is given, and the proposed method shows good result.
Seismic attributes are quantitative measure of the characteristics of a seismic trace over specific intervals. They can provide as much information as possible for us to integrate the subsurface structure and to predict the hydrocarbon. They are commonly used for prospect identification and risking, hydrocarbon play evaluation, reservoir characterization, etc. One advantage of the seismic attributes is that it can predict at and away from wells while still honoring well data and most times, predictions are more detailed than simply interpolating well data. There are now a lot of distinct seismic attributes calculated both from seismic data and their transforms. How to use these attributes to integrate the subsurface structure and predict the reservoir is a problem which we should face. Neural network methods have been developed for reservoir prediction using seismic attributes, Cai et al. (1993) used BP neural networks and Xu et al. (1998) used SOM neural networks to predict reservoir. Support Vector Machine (SVM), first introduced by Vapnik (1995), is a set of related supervised learning methods for classification and regression. SVM gives the best generalization accuracy classifier by maximizing the margin between two classes in the feature space. Because of this, it has been applied widely in many fields such as classification, regression, face recognition and time series prediction. SVR (Support Vector Regression) is an application of SVM for function regression. In this paper, we propose an attributes selection method based on SVM and then use SVR to approach the hydrocarbon reservoir using the chosen attributes. First we make a selection of the attributes to get the attributes which give more contribution to the result and eliminate the useless attributes in order to reduce the empirical risk, and then we will use the chosen attributes for regression.
In pattern recognition, features unrelated to results have a bad effect on the generation performance of classifier. So it’s necessary for us to have a selection among all the attributes before we design the classifier. And this is the same with regression. We can get related attributes so as to minimize the computation and to increase the efficiency while decrease the empirical risk. SVM recursive feature eliminating (RFE) algorithm (Guyon et al. 2002) is an efficient linear attributes ranking method and has been used in many fields. However, this method will take a lot of iterations to give a ranked list of features and this will cost much time.
