From Narrow-Band to Wide-Band Frequency Spectrum: U-Net Network for Seismic Data Resolution Enhancement

High-resolution seismic data processing is critical for energy and resource exploration. Recent advances in deep learning have demonstrated the potential of neural networks in enhancing seismic data resolution. Existing deep learning methods for seismic data can be divided into well-constrained and well-independent approaches. The former relies on well log data, while the latter often faces challenges in maintaining consistency with well data. We propose a U-Net-based approach to predict the high- and low-frequency components of seismic data without relying on well data. Our method uses mid-frequency components with a high signal-to-noise ratio (SNR) as a prior constraint for creating synthetic training datasets. The deep learning model is trained to predict both high- and low-frequency components, thereby broadening the effective frequency range of seismic data. Initially, a theoretical model is employed to validate the predictability of different frequency components. Subsequently, numerical experiments with the Marmousi2 model demonstrate that joint prediction of high and low frequencies yields the highest accuracy. When applied to field seismic data, the proposed method outperforms traditional deconvolution and other deep learning techniques, effectively generating high-resolution seismic data in well-independent scenarios. These results indicate that our approach offers a robust and adaptable solution for enhancing seismic data resolution, which is essential for advancing seismic data-processing technologies.