First Arrival Picking And Inversion Of Exploration Data Based On U-MLP Network

In seismic exploration,first-arrival picking is an important step for time-lapse tomography in seismic data processing,and the accuracy of first-arrival picking also seriously affects the results of inversion,so how to improve the accuracy of first-arrival picking is extremely important for travel time inversion.With the deepening of geological exploration research,the traditional first-arrival picking method requires more manual intervention and calculation effort,which cannot adapt to the growth of modern seismic exploration data scale and the complexity of tasks.At the same time,the traditional first-arrival picking method is prone to the problems of large picking error and low generalization when dealing with low signal-to-noise ratio and lack of seismic data.In order to improve the accuracy and efficiency of first-arrival picking,this thesis takes the first-arrival picking task as a pixel-level classification task in image segmentation,i.e.the image segmentation task,and combines deep learning techniques to study the initial arrival picking task in seismic exploration,the specific work is as follows:Part 1 proposed an improved first-arrival picking method based on the U-Net network.Because the U-Net model using the cross-entropy loss function tends to favor higher-frequency categories when the data category is unbalanced,it is more inclined to appear categories with higher frequency,thus resulting in large errors.In this thesis,a U-Net network model based on WLS loss function is proposed and applied to the first-arrival picking task,which uses a linear combination of weighted cross-entropy and Lovasz-Softmax loss function to solve the problem of data category unbalance,and the Jaccard index is also optimized.The thesis performed comparative experiments based on noisy synthetic datasets,and the results proved that the U-Net model using the WLS loss function has higher picking accuracy than that using the cross-entropy loss function.Then,by constructing a twolayer velocity model,the results picked up by the model are inverted and applied,and the inversion results show that the initial to inversion images picked up by the U-Net model based on WLS loss function is more clearer.Part 2 proposed a first-arrival picking method based on the U-MLP network.This thesis introduced residual connections into the U-Net network to reduce semantic errors and improve the segmentation accuracy during feature fusion.The thesis introduced a Marked MLP module into the U-Net to replace the convolutional blocks in the convolution stage,as well as changing the number of input channels during the input MLP module process,to reduce number of parameters and computational complexity,to make the model more focused on learning local features.The thesis conducted comparative experiments based on noisy synthetic datasets by applying the U-MLP model,and the results show that this model further improved the accuracy of first-arrival picking.By constructing three-layer velocity models,the picking results are applied to the inversion,and the inversion results show that the U-MLP model is more suitable for the inversion of complex velocity models,and the inversion image is clearer than those produced by the U-Net model.Part 3 designed and realized a first-arrival picking system based on the U-MLP model.The system is developed by using Py Qt5 as the front-end development framework,the back-end logic is implemented by involving standard Python modules and writing custom modules,including user login,model selection,data processing,first-arrival picking,and saving the picking results.To make user clearly observe the picking results,the system also supports data visualization before and after picking.The inversion application verification of the initial solstice picked up by the system is verified by the self-compiled inversion script,and the results show that the initial solstice picked up by the system can be well inverted speed model,which reflects the practicability of the system...