Intelligent seismic stratigraphic modeling using temporal convolutional network

Abstract

Deep learning (DL) has stimulated the intelligent processing and interpretation of seismic big data for oil and gas exploration in the past decade. In the applications to seismic stratigraphic modeling, most DL algorithms utilize a combination of ‘convolution + pooling’ to extract spatial features from seismic volume or images, which, however, may lead to a loss of the fine-grained information endowed in seismic time-series records inevitably. Here, we proposed a novel scheme for intelligent seismic stratigraphic interpretation based on temporal convolution network (TCN). The purpose of using TCN is to implement a trace-by-trace segmentation of lithologies by extracting spatio-temporal correlation of seismic record from a small amount of well log labels or manually interpreted geological profiles, thereby building a 3-D subsurface stratigraphic model. The proposed TCN model for seismic stratigraphic interpretation is trained, validated, and tested using the Netherlands F3 seismic dataset. The results suggest that, with a limited number of either manual interpretation profiles or well logs information, the proposed TCN algorithm can produce a highly accurate and laterally continuous 3-D stratigraphic model. Specifically, the mean pixel accuracy of prediction results of the above two schemes (compared to the real model) has achieved 98.23% and 96.09%, respectively. In addition, comparing TCN to the widely used encoder-decoder convolutional neural networks for seismic strata interpretation shows the outperformance of TCN scheme in terms of training time, model size, and prediction accuracy, while also revealing that the proposed TCN model can well identify thin strata in complex seismic data.