一种改进STGCN的深地时空域地震子波提取方法

An improved STGCN method for extracting deep-zone seismic waveletin spatiotemporal domain

地震子波的准确提取可有效提高全波形反演和偏移成像等方法的准确性,对储层预测和油气分析具有重要意义。由于深层能量衰减和复杂地质构造,地震子波不仅具有时变特性,同时也具有不可忽略的空变特性。而传统时变子波提取方法仅通过单道地震记录提取时变子波,忽略了多道地震记录之间子波的空间变化。同时,传统时空域子波提取方法,如经验模态分解(EMD)方法,对测井资料等先验信息依赖程度较高,实际应用范围受限。深度学习为时空域子波提取提供了新的思路,针对以上问题,提出了一种改进时空图卷积神经网络(STGCN)的时空域子波提取方法。首先,根据目标区地震数据分布特征与非平稳性质,建立以非平稳地震剖面为输入,时空域子波为标签的合成训练数据,再利用传统EMD时变子波提取方法逐道提取目标区子波,有针对性地构建以目标区地震剖面为输入,目标区时空域子波为标签的实际训练数据。最后,利用两种训练数据对改进后的STGCN进行训练,使其能够融合提取的子波时空特征,从而实现目标区时空域子波的有效提取。合成数据和实际地震数据的处理结果表明,该方法对于深地时空域子波的提取有效且准确,相较于传统方法更具优越性,具有较好的实际应用价值。

An accurate seismic wavelet is crucial to full waveform inversion,migration,reservoir prediction,and hydrocarbon detection.Due to deep energy attenuation and geological complexities,seismic wavelets are time-variant and space-variant.Traditional time-varying wavelet extraction methods are single-trace approaches disregarding trace-to-trace wavelet variation.Moreover,traditional spatiotemporal wavelet extraction methods heavily rely on such prior information as log data,which limits their practical application.To address aforementioned issues,an improved method called Spatiotemporal Graph Convolutional Neural Network(STGCN)is proposed for extracting spatiotemporal wavelets.Firstly,based on the distribution characteristics and non-stationary nature of seismic data in the target area,synthetic training data is generated with non-stationary seismic data as inputs and spatiotemporal wavelets as labels.Subsequently,the traditional empirical mode decomposition(EMD)method is used to extract wavelets on a trace-by-trace basis in the target area,and actual training data is constructed with seismic data as inputs and spatiotemporal wavelets as labels in the target area.The improved STGCN is trained using these two types of training data to extract the spatiotemporal features of wavelets in the target area effectively.The processing results of both synthetic data and actual data demonstrate the effectiveness and high accuracy of the proposed method in extracting deep spatiotemporal wavelets.Compared to traditional methods,the proposed method exhibits superior performance and significant practical value.