Automatic Seismic Travel-Time Picking Using Transfer Learning

Estimating seismic wave velocity in subsurface formations is one of the key issues in seismic oil and gas explora tion.Seismic tomography is an important research method for detecting underground velocity structures.The key to tomography is to obtain accurate reflection event travel-time data from pre-stack gathers.The difficulties in picking up the reflection events of pre-stack gathers are:(1)because of the huge amount of data to be processed,an automatic picking method must be adopted;(2)because the pre-stack gathers are single-track recordings with weak signals,the events are ambiguous,automatic tracking of the events that are very easy to cross;(3)Because of the low signal-to-noise ratio,it is very difficult to automatically identify the starting time of the reflection wave at a specific layer in the continuous reflection signal sequence;(4)The reflection events are dense and there are cross-merging,multiples,and noise interference,which make it more difficult to pick up multiple events simultaneously.In recent years,the application effect of deep learning in many fields such as computer vision has been outstanding.It can automatically extract nonlinear,abstract,and complex features from massive data,making it a new engine for seismic exploration data processing.However,the limited seismic label data becomes a major bottleneck for training deep learning models.To solve the problem of picking up reflection events in pre-stack seismic gathers,based on convolutional neural network and transfer learning,this paper studies the automatic picking method of multiple reflection events in pre-stack seismic gathers.Firstly,the research is carried out from the relatively easy first-arrival picking problem.We use time-frequency transform to enhance the first-arrival feature and combine transfer learning to develop the first-arrival picking method.Then we introduce a two-branch network to combine time-frequency features and time-space domain features to develop the automatic picking method of multiple reflection events of pre-stack seismic gathers.Finally,the research is extended to the field of automatic interpretation of seismic horizons,and the automatic tracking method of multiple horizons of 3D seismic data volume is developed.The content involves:(1)The presence of the weak first arrival,low signal-to-noise ratio,and phase change makes the first arrival picking difficult,so this paper proposes an automatic first arrival travel-time picking method combining time-frequency transform,convolutional neural network,and transfer learning.We use the continuous wavelet transform to enhance the abrupt changes of amplitude,phase,and frequency associated with first arrivals,rather than any single time or frequency domain features.We consider the first arrival picking a three-class classification problem and use the convolutional neural networks to classify seismic waves as noise,first-break,and post-first-break.We transfer the pre-trained model on the large-scale labeled natural images to the seismic wave classification model to solve the bottleneck of a scarcity of seismic label data.The co mparison of the methods has verified the effectiveness of the method.(2)To solve the problem of picking multiple reflection events in the pre-stack data,we introduce a two-branch convolutional neural network to jointly use the time-frequency images and seismic images to simultaneously extract the time-frequency domain and time-space domain features for feature fusion.Then,the mapping relationship between features and output categories is established for the classification.We pick the travel-time of multiple reflection events of pre-stack data from the classification probability.Further,under the small training data set,transfer learning is used to transfer knowledge between seismic data sets in different geographical regions.The model trained on la nd pre-stack seismic data is used as the initial model and then transferred to the marine seismic data classification model for fine-tuning.The fine-tuned model is then used for offshore pre-stack seismic data prediction and reflection event travel-time picking.The picking results of synthetic seismic data and field data demonstrate the effectiveness of the method.(3)Aiming at the problem of automatic horizon tracking in 3D seismic data,a simultaneous tracking method for multiple horizons in 3D seismic data volume is developed based on a two-branch network and transfer learning method.It is usually necessary to train a specific model for a specific data set,which brings about the problem of high cost.The two-branch network model trained on a 3D seis mic data volume is used as the initial model,and the model parameters are transferred to the classification model of another data set.After fine-tuning,the fine-tuned model is applied directly to the new dataset.The picking results of field data verifies the effectiveness of the simultaneous automatic tracking method for multiple horizons of the 3D data volume and the feasibility of transfer learning between datasets in different areas.(4)Intelligent software for seismic wave travel-time picking is developed for practical application.The software provides the functions of first break picking and seismic reflection event tracking,including a transfer learning algorithm and a two-branch network algorithm.It supports the transferring,training,and app lication of the deep learning model.The software also allows automatic travel-time picking in a simple,easy,and user-friendly way.