Neural Networks versus Statistics: a Comparing Study of Their Classification Performance on Well Log Data

Recently, Neural Networks became a popular method in geosciences in the context of pattern recognition problems. They can be used for classification of well logs, for image processing, for anomaly detection and similar problems. In the past, such problems were mainly solved by statistical approaches. However parametric statistical classification methods suffer from strong assumptions and nonparametric one’s are unsuitable for small data sets. The hypothesis is, that Neural Networks can provide a useful technique, if statistic classification methods fail. Some theoretical considerations can support this suggestion. A standard layered Neural Network type with a linear accumulation and a sigmoid transfer function has been used to classify geophysical well logs into lithological sections. The selected data set consists of four geophysical logs (gamma ray, electrical resistivity, density, neutron neutron) and the known lithology of the wells. The available data has been divided into a training and a test data set. On the basis of the training data the networks have built an universal classification rule. The performance has been evaluated on the test data and compared versus a number of statistical classification methods: Linear and Quadratic Discriminant Analysis, Discriminant Analysis with nonparametric density estimators and knearest-neighbour classification algorithms. An account is given of the comparison results.