基于深度学习的地震速度谱自动拾取研究

In conventional seismic processing workflow, the stacking velocity is commonly picked up manually. Due to the continuous increase of seismic data, especially three-dimensional seismic data, it takes a lot of time and energy to manually pick up seismic velocity spectrum. Because velocity picking requires professionals with rich seismic processing experience, it is very dependent on the seismic processing personnel. Experience and subjectivity may result in unnecessary human error for beginners. In order to solve the problems that may arise in the traditional seismic velocity spectrum picking workflow, this paper proposes an automatic picking method based on deep learning, which calculates and automatically picks up the stacking velocity. This paper uses computer vision research methods to process the velocity spectrum as an image, and designs a Anchor Free-based FCOS (Fully Convolutional One-Stage Object Detection) neural network model that can be used for speed picking. Thus, the problem of picking up stacking velocity from the velocity spectrum is transformed into the problem of intelligent identification of energy groups. When dealing with the low signal-to-noise ratio work area, the DNN (Deep Neural Networks) model is added to fit the global velocity curve according to the characteristics of the poor focusing characteristics of the velocity spectrum energy cluster. By training the model in this paper, the superposition velocity based on energy group in the input velocity spectrum can be automatically picked up, and the "time-velocity" pair sequence containing the survey line number and track set number can be output. The test results of Marmousi model data and actual work area data show that the automatic seismic velocity spectrum picking model designed in this paper has high accuracy and strong robustness, which effectively relieves the burden of manual picking and significantly improves the efficiency while ensuring the velocity picking accuracy.