Application of domain-adaptive network with data augmentation in lithology identification of buried hill igneous rocks

Lithology identification is crucial in oil and gas exploration and development. Although traditional logging techniques and existing intelligent identification technologies have achieved significant progress in identifying lithology within conventional reservoirs, challenges persist in recognizing buried hill igneous rocks. Traditional logging methods rely on empirical rules and static models, which are inadequate for complex geological environments — particularly when severe overlap of logging response values occurs among different lithologies — resulting in poor identification capabilities. While intelligent identification technologies can enhance recognition accuracy by learning complex nonlinear features, they still face issues of insufficient generalization ability and model bias when dealing with data imbalance, feature crossover, and significant data distribution shifts between different wells. To address these limitations, we propose a data-analytically optimized Domain Adversarial Neural Network (DANN) framework for lithology identification. The main contributions of this paper include: (1) proposing an optimized data augmentation strategy to alleviate problems of data imbalance and feature overlap; (2) introducing an automatic feature weighting mechanism within the DANN framework to effectively tackle challenges associated with multi-source feature fusion and data distribution shifts; and (3) validating the proposed method on a real dataset from buried hill reservoirs in the northern South China Sea. The results demonstrate that, compared with traditional logging lithology identification methods and existing intelligent approaches, the proposed method exhibits superior performance in cross-well lithology identification. Additionally, the optimized data augmentation strategy significantly reduces model bias caused by data imbalance and overlapping logging response features, enhancing the overall accuracy of lithology identification. • A modified Domain-Adversarial Neural Network is proposed, incorporating an automatic feature weighting mechanism and neighborhood adversarial training to address the severe data distribution shift and multi-source feature fusion issues encountered in igneous rock identification. • An optimized data augmentation strategy is introduced to alleviate the problems of feature overlap and data imbalance in logging responses observed in buried hill igneous rocks. • Lithology classification performance in deep learning models significantly improved, with macro-average recall rates increasing by up to 16%. • The method was validated on real-world datasets, demonstrating superior performance in identifying buried hill igneous rocks under severe data drift conditions.