人工智能和视速度约束的地震波初至拾取方法

初至拾取是近地表静校正处理的重要步骤之一。随着采集密度的不断提高，地震数据量不断增加，迫切需要发展新的方法解决大数据量的初至自动拾取问题。传统方法通过人工交互拾取和质量控制，在面对庞大数据量的高密度数据时效率很低，而基于深度学习的初至自动拾取方法效率较高。在用于初至自动拾取的各种深度学习算法中，全卷积神经网络（FCN）在语义分割方面有突出优势，它可以处理不同大小的地震道集的数据，并且可以进行高分辨率的像素分类，但是这种方法存在定位精度不足的缺点。U-Net结构是FCN的一种变体，凭借较高精度和易于实现的特点，可以较好解决初至拾取问题，但在数据信噪比较低的情况下准确度会下降。为了解决以上问题，提出四个关键技术点：采用处理流程对振幅进行平衡以提高预测精度；比较U-Net与三种不同复杂度的U-Net变体（UNet++、Wide U-Net和Attention U-Net），从不同角度解决分割问题；选取合适的超参数优化网络；通过视速度约束提高分割图像精度。结果表明U-Net获得了更高的精度和效率，并在陆地地震数据应用中取得了较好的效果。

First break picking method based on artificial intelligence and apparent velocity constraint

Picking seismic first breaks is an important step for correcting near-surface long-wavelength static anomalies.Nowadays,dense acquisition brings exponentially increasing seismic data, so that it is necessary to find a new method to pick first breaks.Conventional methods rely on manual picking and quality control, which is inefficient for large datasets.Compared with conventional methods, deep learning can greatly improve picking efficiency.Among the deep learning algorithms for picking first breaks, Fully Convolutional Networks (FCNs) have outstanding advantages in semantic segmentation, they can process data with variable sizes and perform high-resolution pixel classification.However, such segmentation has shortcomings in locating accuracy.U-Net is a variant of FCN that can solve the problem of first break picking.Although it is characterized by easy implementation,the accuracy decreases when the signal-to-noise ratio is low.In order to eliminate the limitation, this paper proposes four key points:(1) Design a workflow to balance the amplitude of samples,thus improving the network accuracy; (2) Compare three state-of-the-art U-Net variants with varying complexity, including Wide U-Net, UNet++, and Attention U-Net; (3) Optimize the network's hyperparameters with categorical loss and improved activation functions;(4) Use apparent velocity to constrain and improve the segmentation accuracy. Comparison of U-Net and its variants with different complexity has shown that U-Net has the best accuracy and efficiency. Finally, the results over a land dataset are promising.