Prior-knowledge constrained seismic attributes fusion with unsupervised learning for channel sandbody delineation

Channel sandbody delineation is an essential task in reservoir characterization. Multiple seismic attributes clustering methods have shown great potential in automatic channel sandbody delineation. However, the results of the multiple attributes clustering methods may not correspond to sedimentary facies or geological bodies, since prior geological knowledge is not incorporated into the clustering process. In addition, most clustering methods often adopt a fixed attribute combination for interpreting all sedimentary facies, without considering that different facies may be sensitive to different combinations of attributes. To address these issues, we propose the prior-knowledge constrained deep adversarial multi-view clustering (PDAMC) model to fuse seismic attributes for channel sandbody delineation. Firstly, we select control points for indicating target sedimentary facies from the representative time slice and use them as prior geological knowledge. We then select attribute combinations with the highest sensitivity for each target sedimentary facies using Shapley Additive exPlanation (SHAP) and correlation analysis (CA) constrained by control points. The selected attribute combinations for different facies are then fed into the PDAMC model, where control points are regarded as prior geological knowledge to constrain the clustering process. The prior-knowledge constrained loss is also used to guide the latent representation toward the clustering centers and thereby enhance the clustering performance of the PDAMC model. We next generate probability maps for all facies by calculating distances between seismic samples and the clustering centroids of different facies. Finally, we integrate probability maps of different facies using the red-green-blue (RGB) blending strategy to delineate the channel sandbody. Experiments on the Parihaka survey and offshore Louisiana survey demonstrate that the proposed method produces more continuous and accurate channel sandbody delineation results than comparative clustering methods.