AUTHOR=Zhao Ruipu , Zeng Lili , Fu Chendong , Zhao Xiaoqing 

TITLE=Subdivision of river channel sand micro-scale facies with feature attention spatio-temporal network

JOURNAL=Frontiers in Earth Science

VOLUME=Volume 13 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/earth-science/articles/10.3389/feart.2025.1542579

DOI=10.3389/feart.2025.1542579

ISSN=2296-6463

ABSTRACT=IntroductionSedimentary micro-scale facies research is essential for characterizing the lateral and vertical evolutionary patterns and contact relationships within sedimentary facies. This is critical for the redevelopment of high-water-cut oil reservoirs. The complexity of river channel sands, including their horizontal and vertical heterogeneity, well connectivity, and the effectiveness of water injection, necessitates a more refined subdivision of sedimentary facies. Traditional manual identification methods are labor-intensive and prone to subjectivity, highlighting the need for a more automated and precise solution.MethodsThis paper integrates well-logging sedimentology with statistical theory, selecting multiple reservoir and logging parameters to establish a new classification standard for river channel sand sedimentary micro-scale facies. Based on deep learning techniques, we propose a network that combines feature attention and spatio-temporal feature extraction. The feature attention module dynamically assigns weights to logging parameters based on their correlation with the target classification, enhancing the contribution of key parameters to the classification task. Meanwhile, the spatio-temporal feature extraction module fully leverages spatial and sequential information from the logging data, enabling precise identification of river channel sand sedimentary micro-scale facies.ResultsThis method, applied to a real-world oilfield for residual oil development, subdivides deltaic river channel sand sedimentary micro-scale facies into four distinct types. It improves overall accuracy by 8% compared to traditional CNN models and significantly outperforms existing machine learning methods. Notably, the method achieves 100% classification accuracy for certain micro-facies categories, with an overall classification accuracy of 94.9%, demonstrating its superior performance and potential for application in complex sedimentary environments.DiscussionThis approach not only enhances the accuracy of sedimentary micro-scale facies classification but also offers a new framework for analyzing the connectivity between injection and production well groups. The integration of spatio-temporal feature extraction with feature attention significantly improves model performance, especially in the complex, heterogeneous environments typical of river channel sands. This method represents a substantial improvement over traditional models and has broad applicability in the field of reservoir management.