Random Forests lithology prediction method for imbalanced data sets

For the lithology prediction method depending on a supervised machine learning classifier, if the data set has too few samples of target lithology while too many samples of non-target lithology, the classifier trained on this imbalanced data set will cause the prediction results be biased toward the non-target lithology, resulting in poor prediction accuracy of target lithology. With regard to this problem, a Random Forests lithology prediction method for imbalanced data sets is proposed. Firstly, a lithology data set is constructed with lithological logging data as sample labels and seismic attributes and elastic parameters of rock at the uphole trace as sample features. Secondly, the NM-SMOTE algorithm integrating near miss (NM) and synthetic minority over-sampling technique (SMOTE) is employed to balance the lithology data set. Then a Random Forests classifier is trained on the balanced data set to build a nonlinear relationship of lithology with various seismic attributes and elastic parameters. Finally, the seismic attributes and elastic parameters of the target explorato-ry area are input into the Random Forests classifier which will predict lithology according to the above nonlinear relationship obtained during training. The actual data test results demonstrate that too many samples of non-target lithology will affect the prediction accuracy of the Random Forests classifier, and the prediction accuracy of lithology is only 38%. After the training data set is balanced with the NM-SMOTE algorithm, the prediction accuracy of lithology rises up to 83%, and a data volume of lithology is obtained, which is more consistent with seismic data. © 2021, Editorial Department OIL GEOPHYSICAL PROSPECTING. All right reserved.