Seismic Impedance Inversion via Combining Convolutional Neural Network and Geostatistics: An Example from Songliao Basin, China

Abstract

Summary Deep learning methods have been preliminarily applied in several research fields, such as computer vision and natural language processing. Convolutional neural network (CNN) is a kind of deep neural networks, which has the strong abilities of feature extraction and structure representation. Therefore, CNN can learn nonlinear relationship from labelled data and then build the nonlinear model. So, we propose to utilize CNN to do the seismic impedance inversion. However, the validity of CNN is guaranteed by the large number of labelled data. In seismic exploration field, there are usually a limited number of drilled wells that can serve as labelled data. Thus, we propose to utilize geostatistics to augment labelled data according to geological knowledge and rock physics. In order to mitigate CNN’s dependence of labelled data, we propose a framework for CNN’s training, introducing unlabeled data (seismic data) into the loss function. The proposed seismic impedance inversion via combination of CNN and geostatistics is much less dependent on the large number of labelled data and can obtain reasonable inverted results. Real data application verifies the effectiveness and accuracy of the method which can meet the demand of thin interbedded reservoirs identification in Songliao Basin, China.