Cross-well Lithology Identification based on Dynamic Adversarial Adaptation

Lithology identification is a fundamental task in geo-logical work such as stratigraphic correlation, reservoir zonation, and sedimentation simulation. Traditional lithology identification methods require too much manual labor and are inefficient. The shifts in data distribution between differ wells make it difficult to apply the same logging lithology identification method from well to well. In this paper, we study cross-well lithology identification by unsupervised domain adaptation. The features of the logging curve in the depth domain are extracted by semantic segmentation to achieve intensive lithology prediction for each depth point. An adversarial learning strategy is introduced to the cross-well lithology classification task to reduce the difference in distribution between two well features. The proposed method outperforms the baseline in four cross-well lithology identification tasks, demonstrating the efficiency of our method for cross-well lithology identification.