Towards Automated Facies Recognition Using Digital Outcrop Models and Machine Learning

Facies models are an essential tool for characterizing depositional environments in modern and ancient settings and are commonly used to predict spatial facies variability and parameterize reservoir models, particularly in data-poor areas. However, these models generally lack quantitative data (e.g., dimensions, geometries, lithology proportions) and tend to be constructed from end-member scenarios, creating subjective and complex models that are not universally appropriate. This study utilizes supervised and unsupervised machinelearning algorithms to classify lithology and facies from 3D outcrop models collected using drone-derived photogrammetry to build and improve facies models, making them more reproducible and less biased. High-resolution point clouds of siliciclastic turbidite outcrops form the dataset for this study. Rather than labeling subjectively-defined facies (e.g., interbedded sandstone-mudstone), we chose more objectively-defined labels that are consistent with basic lithologies (e.g., sand, mud), and these lithology labels were defined at the centimeter scale. Features (e.g., color, texture, slope, geometry) were extracted from the point clouds using multiple search radii. These features are inputs for Random Forest, K-Nearest Neighbors, and Support Vector Machine algorithms used to classify the lithology of the point cloud. Preliminary results indicate accuracies of 90% using these algorithms for lithology classification. Current work is focused on clustering lithologies into objectively-defined groups (i.e., facies), that in conjunction with field observations (e.g., stratigraphic surfaces, sedimentary structures, ichnofossils), can be used to generate facies models. This clustering utilizes an anisotropic approach that honors the stratigraphic layering and lateral variability present in the point clouds. The goal of this study is to provide a rapid workflow for outcrop point cloud analysis that will generate quantitative lithology and facies data, thereby decreasing the time necessary to build accurate models for subsurface characterization of earth resources (e.g., hydrocarbons, CO2, water).