AI Insight
This study developed an improved deep learning model for lithology identification by integrating the CBAM attention mechanism into a ResNet18 neural network architecture. After systematically comparing five different attention mechanisms, the CBAM-enhanced model achieved 86.36% accuracy in identifying rock types from grayscale images when optimized with specific hyperparameters (learning rate 0.0001, batch size 16). The approach addresses limitations of traditional lithology identification methods, which are typically inefficient, expensive, and heavily dependent on expert knowledge.
Why it matters
Accurate lithology identification is essential for geological exploration, petroleum extraction planning, and construction safety assessment. This automated deep learning approach could reduce costs, improve efficiency, and decrease reliance on specialized expertise in multiple engineering and resource exploration applications.
Understand the Science
by Yan Zhang, Fan Zhang, Jiaming Li, Lantao Huang, Ruifeng Li
Lithology identification is vital for improving geological exploration accuracy, optimizing petroleum exploration schemes, and ensuring construction engineering safety. Traditional identification methods suffer from low efficiency, high cost, and heavy dependence on domain expertise. This study proposes a lithology identification approach based on the ResNet18 backbone, in which five lightweight attention mechanisms, namely CBAM, channel-only attention, ECA, SE, and spatial-only attention, are systematically compared, and the best-performing CBAM module is embedded to strengthen feature extraction and discrimination capacity. The Open-Source Computer Vision (OpenCV) library is adopted for dataset preprocessing, where raw images are converted to grayscale to suppress interference from irrelevant information. Comparative experiments demonstrate that CBAM outperforms the other evaluated attention modules and is therefore incorporated into the final network architecture. A hyperparameter sensitivity analysis of the CBAM-based model demonstrates that the model performance varies with different combinations of learning rate and batch size. When trained with a learning rate of 0.0001 and a batch size of 16, the model achieves a lithology identification accuracy of 86.36%, substantially improving overall recognition performance. By integrating CBAM into ResNet18, the proposed method enables accurate recognition to support multi-field engineering applications, reduce risks, and promote intelligent precision.