Enhancing Multiattribute Seismic Facies Analysis: The Effects of Dimension Reduction on Unsupervised Clustering

Abstract The integration of machine learning (ML) into 3D seismic workflows has shown improved geologic interpretability of deepwater channels in the Taranaki Basin, New Zealand. The integration of two novel dimension reduction (DR) techniques, kernel principal component analysis (K-PCA) and uniform manifold approximation and projection (UMAP), has shown improved interpretability for these channel systems. K-PCA yielded higher-resolution seismic images for identifying internal channel architectural elements (AE), whereas UMAP demonstrated improved resolution for identifying lateral AE when compared to an established DR technique, principal component analysis (PCA). Building on these findings, we shift our focus to determine how the implementation of dimension reduction (DR) on seismic attributes before conducting unsupervised clustering affects cluster performance. Our results demonstrate that pre-DR unsupervised k-means clustering produces non-unique clusters that exhibit low statistical performance, while post-DR clustering with K-PCA and UMAP shows noticeable improvement in both statistical metrics and visual separation of clusters. The interpretation of cluster performance requires careful consideration of dimensional perspective; data that appears highly overlapped in 2D visualizations often reveals distinct cluster separation and better cohesion when examined in 3D space. We propose a balanced evaluation framework that integrates both quantitative metrics (silhouette score and Davies-Bouldin Index) with a qualitative geologic assessment to ensure meaningful interpretation. We aim to present a workflow for adopting and utilizing advanced statistical methods from a geoscience perspective. As ML adoption grows within the field, workflows and interpretational considerations presented by geoscientists for geoscientists become increasingly valuable.