Automatic Recognition Method of Low SNR Regions Based on U-net

Abstract

The SNR (signal-to-noise ratio, SNR) attribute of seismic data is often used to evaluate the processing effect and help finding special geological bodies. Most of the SNR extraction methods need manual analysis to determine the location and time window, which has low precision and long consumption. In addition, the existing methods cannot accurately identify the boundary of low SNR region. This paper presents an automatic recognition method of low SNR regions of seismic data, which uses U-net (U-shaped fully convolutional network) to detect the boundaries of low SNR regions. The core of the method is to convert the recognition of SNR distribution into an image segmentation problem. We gives the specific process, takes the theoretical simulation data with different SNR as samples for network training, and uses the well trained model to segment seismic data with different SNR. The practical application shows the method has strong generalization ability, significantly improves the accuracy and efficiency of SNR recognition, and greatly reduces the workload of manual analysis.