Effective Frequency Band Driven Limited-Sample Deep Learning High-Resolution Seismic Processing Method

The popular deep learning high-resolution seismic processing methods face great challenges in achieving desirable high-SNR (signal-to-noise ratio) high-resolution results of the field seismic data with limited effective frequency band (or dynamic range), due to the limited effective frequency band of seismic data, the limited seismic training sample and generalization performance of the deep learning networks. In this paper, in order to minimize the impacts of the above three issues to obtain desirable high-SNR high-resolution seismic processing result, we proposed an effective frequency band driven limited-sample deep learning high-resolution seismic processing method (ELDHM). ELDHM consists of the effective frequency band driven synthetic seismic training sample generation (ETG), the deep learning high resolution seismic data processing network (DHN) and its loss function, and the effective frequency band driven seismic data optimization (ESO). ETG aims to simultaneously generate the synthetic low-resolution and high-resolution seismic training samples with the same effective frequency band. DHN aims to output the high-SNR high-resolution seismic results. ESO aims to minimize the impacts of the limited effective frequency band of the seismic data and achieve desirable high-SNR high-resolution seismic results without instantaneous phase changing. One synthetic noise-free two-dimensional seismic dataset and one field three-dimensional seismic dataset applications demonstrate that ELDHM achieves high-SNR high-resolution seismic results, and can be used for high-resolution seismic processing.