Background The intelligent identification of seismic facies can significantly improve the efficiency of sedimentary system characterization and hydrocarbon reservoir interpretation. However, influenced by factors such as non-stationary geological bodies, high costs of sample labeling, and limited training samples, conventional methods for intelligent identification are generally insufficient to achieve high identification accuracy and widespread application concurrently.
Advances This study presents a systematic review of three types of technologies for the intelligent identification of seismic facies, namely unsupervised, supervised, and semi-supervised learning, with each type including deep learning methods. The three technological types are comparatively verified using 3D seismic data from a practical survey area in the eastern Amazon region. The results indicate that unsupervised clustering methods, including K-means, self-organizing map (SOM), and generative topographic mapping (GTM), can rapidly reveal the relative distribution patterns of seismic facies. However, these methods are sensitive to parameter setting and initialization while also relying on geological constraints and manual facies assignment typically. These limitations pose challenges in establishing a stable and quantifiable well-tied accuracy benchmarking system. In terms of the supervised learning technology, the validation results based on nine wells indicate that the artificial neural network (ANN), support vector machine (SVM), and random forest (RF) methods yielded macro-averaged accuracies of 79.11%, 81.56%, and 85.78%, respectively. Among the deep learning algorithms in the supervised learning technology, the improved deep dilated convolutional neural network (DDCNN) exhibited an overall accuracy of 91.56%, Cohen's Kappa of 90.35%, and significantly enhanced capacities to characterize the spatial continuity and facies boundaries of seismic profiles. In contrast, semi-supervised learning delivers more pronounced advantages under conditions of limited labels. Specifically, the semi-supervised deep auto-encoder (SSDAE) yielded an overall accuracy of 92.67% and Cohen's Kappa of 91.56%, while the semi-supervised contrastive learning (SSCL) model exhibited an overall accuracy of 91.33% and Cohen's Kappa of 88.97%. Compared to that of the SSDAE, the required labeling ratio of the SSCL model can be reduced from approximately 10%–15% to 3%–4% under comparable accuracy.
Prospects In the future, it is necessary to further integrate geological prior knowledge and interpretable mechanisms, as well as developing self-supervised/contrastive pretraining, the adaptation to cross-data distribution, and the quality control process of uncertainty quantification. These efforts are expected to enhance the robustness and engineering application of the intelligent identification of seismic facies under complex geological conditions. The experimental results further indicate that under typical industrial constraints of high labeling costs and limited samples, semi-supervised learning can maintain high accuracy in the identification of seismic facies using substantially fewer labeled samples. This finding offers a definite and efficient direction for technology selection, thereby supporting the large-scale application of the intelligent identification of seismic facies.
DownLoad: