Machine learning and geostatistical framework for modelling of depth-to-bedrock

Depth-to-bedrock (DTB) models often face challenges related to resolution and accuracy due to insufficient and spatially clustered subsurface data, further compounded by complex anisotropy and non-stationary characteristics. To overcome these difficulties, a novel approach is proposed leveraging information from multiple direct and indirect sources and integrating machine learning for geostatistical prediction of DTB. It was applied to the deglaciated Precambrian Shield landscape of northern Saskatchewan characterized with a relative abundance of mineral exploration drillhole data in addition to the existing geological maps, their interpretations, digital elevation models derivatives and remote sensing data. The workflow begins with preliminary data analysis to estimate statistical stationarity of the study domain and identify sub domains where necessary. A Random Forest classifier was used to model DTB probabilities from indirect data. To ensure local representation of the measured DTB, the probabilities were adjusted by indicator cokriging allowing the reproduction of the spatial structure of direct DTB measurements. The output consists in representative DTB model with local continuous statistical distribution of thickness probabilities with respective minimum, maximum, average, median and quantiles at all locations. The innovative approaches developed in this study are not only limited to northern Canada, they have broader applicability across regions facing the problem of data scarcity worldwide.