Frequency-dependent multiscale network for seismic high-resolution processing

作者：Xu Yanwu, Sanyi Yuan, Huahui Zeng, Junliang Yuan, Yue Yu, Mingxuan Li, Shangxu Wang · 发表于：Geophysics · 年份：2025 · DOI：10.1190/geo2023-0682.1 · 被引用次数：7 · 研究领域：Seismic Imaging and Inversion Techniques、Seismic Waves and Analysis、Seismology and Earthquake Studies

ABSTRACT Seismic high-resolution processing is crucial to enhance the accuracy and reliability of seismic data, particularly in the exploration and development of complex hydrocarbon reservoirs. Conventional high-resolution processing methods, exemplified by sparse spike deconvolution (SSD), typically rely on the assumption of sparse reflectivity. Recent advancements in deep learning have introduced a deep convolutional neural network (DCNN) into high-resolution processing and relaxed it from rigorous physical assumptions. However, DCNN high-resolution processing often suffers from limited generalization capabilities, primarily due to the scarcity of labeled data and variations in data quality. In this study, we identify a frequency learning bias in DCNN high-resolution processing, in which the network initially prioritizes dominant-frequency components before gradually addressing lower and higher frequencies. This bias results in inadequate learning of high-frequency components. We develop a frequency-dependent multiscale network (FMN) informed by the frequency multiscale transformation theory. The FMN aims to improve the frequency learning direction of neural networks, enabling them to simultaneously and efficiently learn information across low-, dominant-, and high-frequency components. Our synthetic and 3D field data experiments demonstrate that our FMN outperforms SSD and DCNN methods, providing superior generalization and higher fidelity in producing high-resolution res...