GSA Connects 2024 Meeting in Anaheim, California

Paper No. 5-3
Presentation Time: 8:45 AM

MODELLING THE THICKNESS AND 3D LITHOLOGICAL PROPERTIES OF SEDIMENTS ABOVE BEDROCK ACROSS ALBERTA USING SPATIALLY AUGMENTED MACHINE LEARNING ALGORITHMS


PAWLEY, Steven, HARTMAN, Gregory, UTTING, Daniel, ATKINSON, Nigel and LIGGETT, Jessica, Alberta Geological Survey, Alberta Energy Regulator, 4999 – 98 Avenue, Edmonton, AB T6B 2X3, Canada

Information on the thickness and composition of sediments above bedrock is fundamental to many geoscience studies and is a prerequisite for regional hydrogeological and geotechnical investigations. However, it can be difficult to estimate these properties across large regions containing varied physiography; traditional methods of interpolating bedrock elevation often underestimate or overestimate bedrock elevation in areas of rugged terrain, and lithological composition is usually inferred from stratigraphic modelling, which is difficult/costly to apply to regions with poorly defined stratigraphy, or where the geology exhibits a large degree of inherent lithological variability.

Here, we describe a machine learning spatial prediction workflow that uses information from traditional digital elevation model derived estimates of terrain morphometry and satellite imagery, augmented with spatial feature engineering techniques to predict sediment thickness and 3D lithological information across Alberta, Canada.

First, compiled measurements of sediment thickness were used with a natural language model lift to predict bedrock depth across all available lithologs (> 300,000). The combined data were used for sediment thickness modelling employing several machine learning algorithms (XGBoost, Random forests, and Cubist) and spatial feature engineering techniques, with the use of spatial lag variables being shown to significantly improve predictive performances. Finally, the properties of the sediments above bedrock were estimated based on the predicted probabilities from a binary coarse/fine-grained 3D machine learning classification model. The results, when evaluated in 3D as iso-surfaces, or sliced at specific depth intervals, for example to show the probability of coarse-grained deposits occurred immediately above the bedrock contact, were successful at delineating the distribution of many major sand and gravel dominant geobodies at provincial and sub-regional scales across Alberta.