Research on seismic impedance inversion method based on pre-training and improved residual network

Abstract

In this paper, we design a novel network architecture based on the principle of seismic impedance inversion. The network combines convolutional neural network and residual network. In seismic impedance inversion, there are usually only a few Wells. However, supervised learning requires a large number of labeled data training networks. In order to solve the above problems, this paper uses two steps to train the network. The first step, the network is pre-trained by using seismic records as the input and a large number of labeled low-frequency information as the output of the network. The second step is to train the network with seismic records as input and a small amount of well data as output of the network. The network can learn the low frequency trend of impedance through pre-training and capture the high frequency characteristics of impedance through re-training. In summary, the network learns the full-band information of impedance through two-step training. We used two typical models with different geological characteristics to prove the effectiveness of the inversion method. We used two typical models of Marmousi II and Overthrust with different geological features to prove the effectiveness of the inversion method.