面向地震波初至智能拾取的超分辨率深度残差方法研究

针对常规语义分割网络在初至拾取中存在的精度低、泛化能力差等问题,基于U-Net网络,结合残差学习模块和亚像素卷积方法,构建了一种超分辨率深度残差网络的初至智能拾取方法(SD-Net)。该方法使用具有跳跃连接的U型网络融合地震数据的多尺度信息,通过端到端的训练方式简化工作。首先,在SD-Net的下采样阶段引入残差学习模块,克服深层网络退化问题,有效提高对地震数据的学习能力;其次,上采样阶段采用亚像素卷积方法,通过卷积和多通道间的像素重组实现特征图超分辨率重建,以更高精度定位初至;另外,利用迁移学习将模型应用于中、低信噪比模拟数据,仅需少量标注数据即可训练得到最优初至拾取模型。实际算例表明：与U-Net方法相比,SD-Net训练效率明显提高;网络模型具有更高准确率和鲁棒性;迁移学习模型预测的结果验证了SD-Net具有较强的泛化能力;该方法在实际生产应用中对实现高效、准确的初至智能拾取具有重要意义。

Research on depth residual method of super-resolution for intelligent seismic wave first arrival pickup

In view of the low precision and poor generalization ability in the first arrival pickup of conventional semantic segmentation networks, an intelligent first arrival pickup method for depth residual network of super-reso-lution (SD-Net) based on a U-Net network, residual learning module, and subpixel convolution method is proposed. This method uses a U-shaped network with a jump connection to fuse multi-scale information of seismic data and simplifies the work through end-to-end training. Firstly, the residual learning module is introduced in the downsampling stage of SD-Net to overcome the deep network degradation problem and effectively improve the learning ability of seismic data. Secondly, the subpixel convolution method is used in the upsampling stage to achieve the super-resolution reconstruction of the feature map through convolution and multi-channel pixel recombination, and the positioning of the first arrival is achieved with higher accuracy. In addition, transfer learning is utilized to apply the model to the simulated data with a medium and low signal-to-noise ratio (SNR), and the optimal first arrival pickup model can be obtained by training only a small amount of labeled data. Practical examples show that the training efficiency of the SD-Net is significantly improved compared with that of the U-Net method. The network model has higher accuracy and robustness. The results predicted by the transfer learning model prove that SD-Net has a strong generalization ability. This method is of great significance for realizing efficient and accurate intelligent first arrival pickup in actual production.