基于混合网络U-SegNet的地震初至自动拾取

传统初至拾取方法拾取效果和效率不能兼顾、算法稳定性差、工业化应用成熟度不高；基于深度学习的初至拾取方法制作标签耗时费力、数据预处理过程繁琐、网络结构过于复杂，导致训练和测试效率较低。为此，将U-Net与SegNet深度学习网络的优点相结合，构建新的混合网络U-SegNet，并基于U-SegNet自动拾取初至。U-SegNet以SegNet结构为基础，通过在解码器网络的反卷积层之前融合跳跃连接信息，提供编码器网络的多尺度信息，以获得更好的性能，并且其上采样操作将U-Net中的反卷积改为反池化，池化索引被传递到上采样层，网络模型收敛更快。因此，U-SegNet网络结构更利于分割背景噪声区域和含噪信号区域，从而提高初至拾取精度。基于U-SegNet的初至自动拾取流程包括制作训练数据集、设计网络模型、训练网络模型、测试网络模型和实际资料应用。测试和应用结果表明，所提方法的初至拾取效率约为某商业软件的2.2倍，且易于工业化应用，具有良好的发展前景。

Automatic picking of seismic first arrivals based on hybrid network U-SegNet

The traditional first arrival picking method cannot take into account picking effect and efficiency,the algorithm stability is poor,and the industrial application has not been very mature.The first arrival picking method based on deep learning is time-consuming and labor-intensive,the process of data preprocessing is cumbersome,and the network structure is too complex,resulting in low training and test efficiency.Combining the advantages of U-Net with those of SegNet,a new hybrid network U-SegNet is constructed,and based on which first arrivals can be picked automatically.Based on the SegNet structure,U-SegNet provides multi-scale information of the encoder network by fusing jump connections information before the deconvolution layer of the decoder network to obtain better performance,and its upsampling operation changes the deconvolution in U-Net to unpooling.Because the pooling index is passed to the upsampling layer,the network model converges faster.Therefore,the U-SegNet network structure is more conducive to segmenting the background noise area and the area where background noise and valid signal overlap,thereby improving the accuracy of first arrival picking.The first arrival automatic picking process based on U-SegNet includes making a training data set,designing a network model,training the network model,testing the network model and applying it to real seismic data.Tests and applications of the U-SegNet model show that the picking efficiency of the proposed method is about 2.2 times that of a commercial software.It is easy to be industrialized and has a good future in large-scale application.