MODELLING DEPTH TO BEDROCK COMBINING MACHINE LEARNING AND GEOSTATISTICS: A PILOT STUDY IN NORTHERN SASKATCHEWAN

Representation of surficial sediment thickness across northern Canada in global datasets lacks resolution and precision. The lack of reliable thickness estimates is a challenge for mineral exploration, permafrost, groundwater, and geotechnical and infrastructure studies. Increasing interest in environmental stewardship for sustainable development under ESG guidelines, and for climate change further highlights the need for, and value of depth to bedrock interface information. Building surficial geology thickness maps is challenging given the sparse and preferentially clustered boreholes that must be combined with interpretative information and indirect measurements. In addition to being accurate, the maps must also provide a level of confidence and uncertainty in the estimates.

A machine learning workflow is proposed to generate thickness maps fusing several sources of information. Direct measurements of the thickness are precise but sparse and preferentially located around mineral rich areas such as faults. Indirect but abundant measurements such as geological maps, topography, remote sensing and geological interpretations data do not provide the actual depth.

Using Random Forest classifiers, indirect measurements are used to model thickness probabilities. These probabilities are adjusted to match ground measurements using a geostatistical technique called indicator cokriging. The results are thickness distributions (min, max, mean, median, quantiles, etc.) at all locations. This work emphasises the need for a dual approach between the integration of data into spatial trends and local anchoring to the ground measurements. Surficial information alone cannot predict thickness with great accuracy except for the outcrops themselves. It can, however, provide trends in thickness. The subsequent geostatistical approach can then integrate these trends with the ground measurements to bolster accuracy. The workflow is applied on a pilot zone in northern Saskatchewan.