Enhancing seismic resolution based on U-Net deep learning network

Deep and ultra-deep reservoirs, unconventional hydrocarbons, and other complex reservoirs are being developed for oil and gas exploration, high-resolution seismic data with a high signal-to-noise ratio is required for accurate reservoir description. The traditional high-resolution processing techniques, such as the inverse Q-filtering technique based on the stratum attenuation model and the convolution model-based technique, are entirely model-dependent. In this study, we built a deep learning network based on the U-net and suggested a processing technique to boost seismic data resolution. We incorporated ResPath structure into the network and employ a weighted MAE and MS-SSIM combination as the loss function, and added a training strategy to the data processing workflow. Finally, the network is validated using both field data, our suggested network can further minimize the loss of low-frequency components during conventional deep learning high-resolution processing, effectively enhancing the ability to perceive low-frequency seismic data components, the signal-to-noise ratio and resolution of seismic data have both been significantly enhanced.