一种基于改进的U-Net网络的初至自动拾取研究

Research on first-break automatic picking based on an improved U-Net network

针对传统的地震波初至拾取方法对低信噪比资料拾取精度较低、算法的鲁棒性较差,以及目前提出的基于深度学习的初至拾取方法制作训练样本耗时耗力、训练样本尺寸太大或网络结构太深导致训练和测试网络模型效率较低等缺点,本文对经典的U-Net网络结构进行了改进,将经典的U-Net网络结构中的跳跃连接改为包含多个卷积块的残差连接,减小了网络结构中融合的两个图像特征的差异,并使用自动拾取的小尺寸训练和测试样本,对本文用于初至拾取的经典U-Net网络模型和改进的U-Net网络模型分别进行了训练和测试.结果表明,改进的U-Net网络模型的训练准确率更高,初至拾取的精度也更高,尤其对低信噪比地震道初至拾取效果较好.

According to the traditional first-break picking method of seismic waves, it has low accuracy in picking low-noise-ratio data, and the algorithm is less robust, and the currently proposed first-break picking method based on deep learning involves the time-consuming and labor-intensive process of making training samples, the size of the training samples is too large, or the network structure is too deep, which results in more efficient training and testing network models, this article improves the classic U-Net network structure by changing the skip connection in the network structure to a residual connection that includes multiple convolution blocks, reducing the difference between the two image features fused in the network structure, and using small size training and test samples that are automatically picked up, the classic U-Net network model and the improved U-Net network model used for the first arrival pick-up in this paper were trained and tested respectively. The results show that the training accuracy of the improved U-Net network model is higher, and the accuracy of first-break picking is higher, especially for low signal to noise ratio seismic traces.