Adaptive Gaussian mixture model and convolution autoencoder clustering for unsupervised seismic waveform analysis

Abstract

Seismic facies analysis can effectively estimate reservoir properties, and seismic waveform clustering is a useful tool for facies analysis. We have developed a deep-learning-based clustering approach called the modified deep convolutional embedded clustering with adaptive Gaussian mixture model (AGMM-MDCEC) for seismic waveform clustering. Trainable feature extraction and clustering layers in AGMM-MDCEC are implemented using neural networks. We fuse the two independent processes of feature extraction and clustering, such that the extracted features are modified simultaneously with the results of clustering. We use a convolutional autoencoder for extracting features from seismic data and to reduce data redundancy in the algorithm. At the same time, weights of clustering network are fine-tuned through iteration to obtain state-of-the-art clustering results. We apply our new classification algorithm to a data volume acquired in western China to map architectural elements of a complex fluvial depositional system. Our method obtains superior results over those provided by traditional K-means, Gaussian mixture model, and some machine-learning methods, and it improves the mapping of the extent of the distributary system.