3-D Basement Relief and Density Inversion Based on EfficientNetV2 Deep Learning Network

• Published in: IEEE transactions on geoscience and remote sensing Vol. 62; pp. 1 - 15
• Main Authors: Zhang, Yu, Xu, Zhengwei, Xian, Minghao, Zhdanov, Michael S., Lai, Changjie, Wang, Rui, Mao, Lifeng, Zhao, Guangdong
• Format: Journal Article
• Language: English
• Published: New York IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Basements; Composite scaling technique; Data models; Deep learning; Density; EfficientNetV2; fused-MBconv; Geology; Gravity; Gravity anomalies; Interfaces; inversion; Lake basins; Natural gas exploration; Noise; Oil and gas exploration; Oil exploration; Parameter identification; Scaling; Three dimensional composites; Three-dimensional displays; Training
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2024.3427711

Gravity interface inversion is a critical technique in delineating the substructure of basins, providing essential technological and data support for oil and gas exploration. Traditional gravity inversion approaches often encounter issues such as suboptimal local solutions and limited resolution. Moreover, conventional deep learning inversion methods typically require extensive time for empirical parameter adjustment, hindering the achievement of optimal training outcomes. By utilizing Bouguer gravity anomaly data, this research pioneers the application of the EfficientNetV2 network in predicting 3-D basement relief interfaces and variations in overburden density. The network employs a composite scaling technique to adaptively adjust its width, depth, and input resolution, thereby identifying the most effective network configuration. Concurrently, the innovative Fused-MBconv convolutional module efficiently achieves superior results with a reduced number of network parameters. Specifically, in the Poyang Lake Basin study in Jiangxi Province, China, the EfficientNetV2 model demonstrated enhanced accuracy in predicting density variations of the basement interface and overlying strata compared to traditional methodologies.