Channel attention-based static-dynamic graph convolutional network for lithology identification with scarce labels

Accurately identifying the complicated lithology distribution of unconventional reservoirs has become an urgent need for oil and gas exploitation. However, the expensive and rare lithology labels limit the precision of current machine learning-based lithology identification methods. Meanwhile, those methods neglect the associated in-formation implicit in adjacent logging samples and underutilize the abundant unlabeled data. In this paper, a channel attention-based static-dynamic graph convolutional network for lithology identification is proposed to solve the problem of scarce labels. To utilize the unlabeled logging data adequately, logging graphs are modeled based on geographic distance and feature similarity to characterize associated information as well as logging features. On the basis, the dynamic similarity graph is introduced to magnify distinctions and reduce connections between samples with different lithologies, improving the quality of logging graphs. Aiming at enhancing the model performance of eliminating redundant and noisy information in logging graphs, the channel attention mechanism is taken into account to realize the fusion and complementation between different associated in-formation. Verification experiments are conducted by utilizing the logging datasets from a shale formation. Results show that the proposed approach has general applicability and achieves high identification accuracy for various strata. Moreover, the proposed approach reduces the cost of drilling coring for lithology identification immensely and can serve as a reference for applying graph theory to solve future reservoir-related issues in petroleum exploitation.