Thickness prediction of seismic multi-attributes sand based on association rules and random forests

Abstract

Seismic attributes analysis technique is an important tool for sand thickness prediction. Due to the varieties of seismic attributes, the best seismic attributes need to be optimized before the seismic attributes analysis technique is applied to reduce the repeatability and redundancy of the attributes. Therefore, we present an improved random forest regression algorithm combined with associate rules for sand thickness prediction (AR-RFR). Although random forest regression algorithm(RFR) is powerful for the problem characterized for nonlinearity and high dimension in reservoir prediction, it cannot solve attribute reduction problems. The associate rules can find the non-linear relationship among the multi-attributes and can reduce some redundant attributes by means of Chi-squared Test. We apply ordinary RFR, AR-RFR and Neural network regression algorithm combined with associate rules(AR-BP) to a synthetic geological model and a real dataset. The results prove that the selection attributes from associate rules is more efficient than that from random forest. Compared to the drilled wells, AR-RFR has higher precision than RFR and AR-BP. And AR-RFR also can improve the lateral distribution of sand bodies. The method proposed is able to choose efficient seismic attributes and improve prediction of sand thickness.

Keywords

• seismic attribute selection
• association rule
• frequent pattern tree
• random forest
• reservoir prediction