Near Surface Seismic Tomography And Machine Learning Applications In Statics Corrections

In seismic exploration,especially in land and shallow marine areas,complex near-surface topography and geological conditions bring great challenges to seismic data processing.Statics correction is a key method to solve the near-surface problem.The long-wavelength statics can be resolved with near-surface velocity inversion,which needs high-quality first-break picking results.The short-wavelength statics is often derived from data-based residual statics correction,in which residual statics solution is generally coupled with stacking velocity analysis and normal moveout correction.However,velocity picking could be time consuming.As the numbers of channels in seismic acquisition systems increase,statics correction seriously affects the efficiency of seismic data processing.In addition,it is very important to obtain an improved first-arrival traveltime tomography result,which can not only provide a better model-based statics solution,but also provide a better initial near-surface model for waveform inversion and depth imaging.As a result,there is a great need and challenge for efficiency and accuracy in solving the near-surface problem of seismic exploration.For first-break picking,many algorithms have been proposed to automate the first-break picking,which,however,remains a challenging problem and still requires significant human efforts.In this thesis,we develop two automatic first-break picking methods based on integrating geophysical methods with machine learning approaches.When there is a small number of labeled samples used for training,we implement a support vector machine method with multi-trace and multi-attribute analysis to improve the automatic picking.we first apply conventional picking methods to compute different properties(e.g.,energy ratio and fractal dimension)of the signal.Considering the waveforms similarities over adjacent traces,we further compute waveform correlation to utilize the property of horizontal consistency.Due to diverse properties,especially horizontal consistency,and large similarity of data in the same project,the machine learning model needs to be trained on only a small amount of data(less than 5%)but can be applied to infer most of the remaining data.We test our method on three real datasets.Compared with the short-and long-window ratio method,our method leads to lower picking error and reciprocal error.In addition,we develop an iterative process that a traditional seismic automated picking method to obtain preliminary first breaks and uses a machine learning method to identify,remove,and fix poor picks based on training big data.We construct a convolutional neural network architecture to help identify poor picks across multiple traces and future refill the picks with the help of the trained model.To allow the trained model applicable to various regions and different datasets,we apply moveout correction with preliminary picks and address the picks in the flattened input.We collect 11,239,800 labeled seismic traces.During the training process,the model's classification accuracy on the training and validation data sets reaches 98.2%and 97.3%,respectively.The prediction results on new 2D and 3D datasets demonstrate the feasibility of our method.Reflection residual statics is often derived from stack power maximization with assumptions valid for NMO and stacking.In this study,we present a deep learning method for deriving surface-consistent reflection residual statics directly from common shot,common receiver,or common midpoint gathers.We augment training data by using only a few gathers but massive possibilities for statics values.Because the magnitude of residual statics is often small,we apply a high-resolution neural network as the backbone to learn detailed features.We also design the head of neural network to allow multi-scale training and multi-scale testing,which enables different acquisition geometries for prediction.Training with synthetics or real data has been tested with real data as input for prediction.In both cases,residual statics helps improve stacked sections.The residual statics solution derived from the trained model is similar to that of stack power maximization.For first-arrival traveltime tomography,different regularization methods or constraints are adopted in the objective function to stabilize the inversion process and ensure more accurate inverted model.Generally,the statics correction derived from the inverted model is used to evaluate different tomography methods.In this thesis,we explicitly apply the optimized statics to constrain the first-arrival traveltime tomography.The optimized statics can be approximately derived by combining long-wavelength statics and residual statics.Compared with Tikhonov regularization,the statics constraint can improve the inverted model and provide better initial near-surface velocity model for waveform inversion and depth imaging.With integrating machine learning and geophysics,our methods can improve the effectiveness and efficiency of first-break picking and residual statics correction in solving the near-surface problem.Based on the statics constraint,we can further improve the first-arrival traveltime tomography.