Application Research Of First-break Picking Based On Full Convolutional Neural Network

First-break,which means the moment when the seismic wave first arrives,is the necessary information for seismic data processing.In recent years,with the complexity of exploration tasks and the expansion of exploration scale,first-break picking work is faced with many new challenges.Among them,the main problems are the lack of seismic records and low signal-to-noise ratio(SNR).For this reason,it is often necessary to introduce additional processing links,such as interpolation or denoising.These links not only increase the consumption of manpower and material resources,but also affect the accuracy of first-break pickup results.Fully convolutional neural network is a deep learning tool widely used in the field of image and speech.Its powerful ability of data information mining and representation can effectively deal with a variety of problems.In this paper,the problem of first-break pickup is transformed into an image semantic segmentation problem by taking first-break as the boundary of background and effective wave.Because of the end-to-end characteristics of deep learning,the method can directly process data with missing channels or low SNR by fine-tuning and discard additional processing links.Through the visualization,analysis and comparison of network intermediate results,this paper shows the network's processing mechanism of first-break picking on the data with traces missing and low SNR.The interpretability of the method is further increased.2273 single shots from actual data were tested by methods in this thesis.The results showed that the classification accuracy of the method was more than 99% on average for the single shots with the proportion of traces missing within 50%.For the single shots processed by adding 20% noise,the method can accurately pick up more than 80% of 2273 single shots,that is,the error of the result is controlled within the allowable manual pick up error range.