MCSEM induction-polarization anomaly identification based on two-scale feature extraction network and XGBoost

Abstract

Due to the presence of induced polarization effect in subsea reservoirs, marine controlled-source electromagnetic (MCSEM) data contain induction response and polarization response. The traditional magnitude versus offset (MVO) curve makes it difficult to manually identify the induction-polarization anomalies contaminated by noise, which leads to the reduction of anomaly resolution and affects the accuracy of data interpretation. Machine learning models possess strong feature extraction and classification ability, which can obtain probabilistic anomaly classification results. Therefore, this study proposes an induction-polarization anomaly identification method based on a two-scale feature extraction network (TFEN) combined with XGBoost algorithm. First, MCSEM induction-polarization theoretical data are calculated using the Cole-Cole model, with random noise added to simulate noisy field data. Then, to effectively fuse muli-scale features while maintaining computational efficiency, a TFEN model is constructed. This model employs long short-term memory and dilated convolution to automatically extract the two-scale nonlinear features from induction-polarization data, followed by feature fusion. Finally, the identification of MCSEM induction-polarization data is realized using the XGBoost. The results show that TFEN-XGBoost achieves the highest anomaly identification accuracy compared with Random Forest, TFEN alone, and XGBoost alone. When the MVO curve fails to distinguish induction-polarization anomalies, the TFEN-XGBoost model achieves a recognition accuracy of 95.89% on theoretical data and 88.93% on noisy data sets. This demonstrates that the combined TFEN-XGBoost model can effectively identify induction-polarization anomalies, providing important technical support for oil resource exploration based on MCSEM.