基于特征强化U⁃Net的地震速度反演方法

Seismic velocity inversion method based on feature enhancement U⁃Net

基于深度神经网络的地震速度反演方法面临的挑战是:时间域地震数据与空间域模型信息间语义映射的弱对应关系导致多解性;神经网络将地震数据映射到速度模型过程中缺少有效引导,易受噪声干扰,影响反演精度。为此,提出一种基于特征强化U‑Net的地震速度反演方法。首先,通过多炮地震数据特征叠加使输入网络的地震时间序列信号与对应速度模型之间的空间关系更加明确;其次,基于多尺度特征融合的思想设计具有不同尺寸卷积核的模块,以增强网络对有效特征的学习能力;然后,利用注意力门引导网络,增强网络重点关注的特征;最后,结合瓶颈残差和预激活的思想,在网络中加入预激活瓶颈残差,避免梯度消失和网络退化。实验表明,该方法在地震速度反演方面具有更高的精度,并在抗噪声测试中效果较好,具有一定的泛化能力。

The challenge faced by seismic velocity inversion methods based on deep neural networks is that the weak semantic mapping correspondence between seismic data in the time domain and model information in the spatial domain leads to a high degree of multiplicity.Additionally,neural networks lack effective guidance in mapping seismic data to velocity models,making them susceptible to noise interference and thus affecting inversion accuracy.Therefore,a seismic velocity inversion method based on feature enhancement U‑Net is proposed.Firstly,by integrating the features of multi‑shot seismic data,the spatial relationship between the seismic time series signal input to the network and the corresponding velocity model becomes more apparent.Subsequently,based on the concept of multi‑scale feature fusion,modules with convolutional kernels of varying sizes are designed to bolster the network’s capacity for learning effective features.Next,attention gates are used to guide the network and enhance the features that the network focuses on.Finally,based on the bottleneck residual and pre‑activation,a pre‑activation bottleneck residual is incorporated into the network,to avoid gradient disappearance and network degradation.The experiment shows that this method has higher accuracy in seismic velocity inversion and performs well in noise testing.It has a certain generalization ability.