Automatic Lithology Identification of Sandstone-type Uranium Deposit in Songliao Basin Based on Ensemble Learning

Abstract

The accurate identification of stratigraphic lithology is closely related to the delineation of sandstone-type uranium deposits. In the face of complex stratigraphic structure, the correct analysis of lithology combination is of great significance to the exploration and anomaly identification of sandstone-type uranium deposits. In uranium exploration, geophysical logging data, as a bridge between the change of geophysical properties and the underground geological environment, is an effective and irreplaceable method to understand the underground rock structure and reservoir characteristics. Conventional lithology identification methods such as crossplot method, probability statistic method, cluster analysis method and conventional machine learning class method have some defects, such as low accuracy, identification efficiency and generalization ability. Ensemble learning is a method of achieving consensus in predictions by integrating significant attributes of two or more models, making the final learning framework more comprehensive than that of a single component model, reducing errors and other factors. Compared with ordinary machine learning algorithms, integrated learning algorithms have more advantages in data processing. In this paper aiming at the problems existing in traditional logging lithology identification methods and machine learning methods, the sandstone-type uranium ore in Songliao basin in north China was taken as the research object, and the original data were analyzed and pretreated. Combined with previous studies, two typical integrated algorithm models (XGBoost and SMOTE random Forest) were used to carry out automatic lithology identification of sandstone-type uranium ore in Songliao basin, and the recognition results of the two integrated algorithm models were compared with K-Nearest Neighbor (KNN), Gradient Boosting Decision Tree (GBDT) and other typical machine learning algorithm models were also compared. The results show that the accuracy of XGBoost and SMOTE stochastic forest integrated algorithm model for lithology identification of sandstone-type uranium ore is above 95%, and the accuracy of KNN model and GBDT model is significantly improved. In order to solve the problem of overfitting in operation, XGBoost algorithm model was used to control the regular term of overfitting and node splitting, and support characteristic multithreading to calculate the gain, which improves the operation efficiency and ensures the reliability of the integrated algorithm model. SMOTE synthetic minority oversampling technique solves the problem of sample data imbalance in the random forest algorithm model. The optimization process based on integrated algorithm model provides a theoretical basis for lithology classification of sandstone-type uranium deposits, and provides technical support for strategic breakthrough in uranium exploration.

Keywords

• xgboost
• smote
• ensemble learning
• sandstone-type uranium deposit
• lithology identification