Neural net ensembles for lithology recognition

Abstract

Lithology recognition is a common task found in the petroleum exploration field. Roughly speaking, it is a problem of classifying rock types, based on core samples obtained from well drilling programs. In this paper we evaluate the performance of different ensemble systems, specially developed for the task of lithology recognition, based on well data from a major petroleum company. Among the procedures for creating committee members we applied Driven Pattern Replication (DPR), Bootstrap and ARC-X4 techniques. With respect to the available combining methods, Averaging, Plurality Voting, Borda Count and Fuzzy Integrals were selected. The paper presents results obtained with ensembles derived from these different methods, evaluating their performance against the single neural network classifier. The results confirm the effectiveness of applying ensembles in real world classification problems. © Springer-Verlag Berlin Heidelberg 2003.