基于全卷积残差收缩网络的地震波阻抗反演

Seismic wave impedance inversion based on the fully convolutional residual shrinkage net

卷积神经网络对地震波阻抗反演已经能取得不错的效果,但反演精度、抗噪声性能有待提高,针对此问题,本文提出了一种基于带逐通道阈值的全卷积残差收缩网络(FCRSN-CW)的地震波阻抗反演方法。该方法首先在残差网络的结构上加入了“注意力机制”和“软阈值化”构成反演网络,然后用波阻抗数据通过正演计算得到合成地震数据集,接着用该数据集训练全卷积残差收缩网络,最后将地震数据输入到训练好的网络中,直接得到反演结果。理论模型反演结果表明,该网络能准确地反演出波阻抗,具有良好的学习能力和抗噪声性能。实测数据反演结果表明,该方法能有效解决地震波阻抗反演问题。

Convolutional neural networks(CNNs)have achieved good results in seismic wave impedance inversion,but the inversion ac-curacy and anti-noise performance need to be improved.Hence,this study proposed a seismic wave impedance inversion method based on the fully convolutional residual shrinkage network with channel-wise thresholds(FCRSN-CW).In this method,the attention mecha-nism and the soft thresholding were first added to the structure of the residual network to form a inversion network.Then,a synthetic seismic dataset was obtained through forward calculation using wave impedance data.Subsequently,the dataset was applied to train the FCRSN-CW.Finally,the seismic data were put into the trained FCRSN-CW to obtain the inversion results directly.The inversion results of the theoretical model show that the FCRSN-CW can accurately invert the wave impedance and possesses satisfactory learning capacity and anti-noise performance.The inversion results of field data demonstrates that the method based on FCRSN-CW can effectively a-chieve seismic wave impedance inversion.