Unsupervised seismic facies classification based on multi-feature fusion autoencoder

Seismic facies classification plays an important role in seismic data interpretation, which is a bridge between seismic data and sedimentary facies. To improve the accuracy of seismic facies classification and reduce the number of manual labels, we propose an unsupervised seismic facies classification method based on the Multi-Feature Fusion Autoencoder (MFAE). At first, the proposed MFAE is generated by the hybride convolution network and variational autoencoder, which extract a large number of latent variables from seismic data. Then, the nonnegative matrix factorization is utilized to implement the principal eigencomponent decomposition and the K-means clustering is also introduced to obtain the results of seismic clustering. Finally, the proposed method is applied to the real data to test the performance. The results reveal that the extracted features of the proposed methods approximately satisfy the Gaussian distribution and the main features after principal eigencomponent decomposition contains the responses of different seismic facies classes. Therefore, the accuracy of the seismic facies classification can be improved. For the Paleogene Shahejie formation in Dongying Sag, the proposed method can predict more clear boundaries of the six seismic facies classes, which is beneficial to demonstrating the evolution of the deltaic sedimentary.