On the capability of support vector machines to classify lithology from well logs

The increasing technical demands placed on models with promising generalization performance to identify lithology from well-log data points has led to search for more efficient methods than conventional statistical methods. Conventional methods such as discriminant analysis recently applied neural networks belong to the general class of Empirical Risk Minimization techniques which aim to minimize training error. On the other hand, methods built on support vector machines (SVMs) are based on the Structural Risk Minimization principle which in turn is based on statistical learning theory. Statistical learning enhancement gives better generalization abilities by minimizing the testing error. In this research, a new modeling framework based on SVMs is described to identify lithfacies from well logs. A SVM classification formulation is presented together with feature selection based on fuzzy theory to identify potential features to enable discriminatory power from well logs. The results demonstrate that SVMs provide an accurate means to identify lithofacies in a heterogeneous sandstone reservoir. The true classification was based on detailed core description of a training well. The SVM-based lithology classifier was compared to discriminant analysis and probabilistic neural networks. The results reveal that the SVM classifier performs the best of the three methods. © 2010 International Association for Mathematical Geology.