Machine Learning Approach for Predicting Depth to Bedrock in Alberta, Canada

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Apr 12 2024

In an article published in the journal Plos One, researchers presented a machine learning approach for predicting the thickness of sediments above bedrock, known as depth to bedrock (DTB), in Alberta, Canada.

Study: Machine Learning Approach for Predicting Depth to Bedrock in Alberta, Canada. Image Credit: Arturo Verea/Shutterstock

They used traditional terrain morphometry and satellite imagery data, augmented with spatial feature engineering, to train several algorithms and spatially lagged DTB estimates. The method significantly improved DTB prediction accuracy in varied terrains and was reliable for provincial-scale mapping. 

Background

The DTB, or the thickness of sediments above the bedrock surface, is a crucial factor in earth and environmental sciences. This measure influences hydrogeological assessments, mineral exploration, and seismic hazard evaluation, among other applications. Traditional DTB mapping often relies on borehole data and spatial interpolation methods, such as kriging and inverse-distance weighting. However, these methods face challenges in areas with rugged terrain or limited data, often producing overly smooth and inaccurate predictions in regions with varied topography.

Geophysical surveys have offered more detailed DTB information but can be expensive and require sufficient borehole data for calibration. Meanwhile, studies using machine learning approaches like neural networks and ensemble tree methods have seen moderate success in DTB prediction, but they have mostly been applied to areas with shallow DTB or regions where bedrock follows the land surface closely.

To bridge these gaps, the present study employed machine learning to improve DTB mapping in Alberta, Canada, using a large dataset of subsurface data and terrain predictors. The researchers supplemented existing bedrock picks with a natural language model to enhance training data coverage. They also incorporated spatial feature engineering strategies and various machine learning algorithms to better account for spatial autocorrelation and data structure, leading to more accurate DTB models and predictions. 

Machine Learning Approaches for Predicting Depth to Bedrock

The authors presented a machine-learning approach to predict the DTB across a province using data from geological boreholes, bedrock outcrops, and pseudo-observations. The research area included various terrains, such as mountains, uplands, and plains, which necessitated the use of advanced modeling techniques to handle diverse spatial patterns.

The researchers utilized different machine learning algorithms, including random forests, extreme gradient boost (XGBoost), and Cubist regression trees, to model the DTB based on a suite of morphometric and spectral features derived from openly available datasets. To account for limited data in certain regions, the researchers employed natural language processing (NLP) to classify water well lithologs as either bedrock or surficial. This classification allowed for the identification of bedrock tops based on predicted litholog intervals.

The study explored various feature engineering approaches, such as incorporating x/y coordinates, using distances to reference locations, and utilizing spatial lag variables, to improve the performance of the machine learning models. Model validation was performed using 10-fold cross-validation and nested three-fold cross-validation to optimize model hyperparameters. The researchers compared different algorithms and spatial feature engineering techniques in four sub-regions with varying geological characteristics.

The optimal machine learning model was then retrained on the full dataset to predict DTB across the province. The authors compared the machine learning predictions with traditional interpolation methods, such as inverse-distance weighted (IDW) and ordinary kriging (OK), to evaluate the quality and geological plausibility of the spatial predictions.

Insights from Bedrock Depth Prediction Models

The researchers assessed lithological and color descriptors commonly associated with bedrock and surficial litholog descriptions, providing insight into common terms such as 'shale' and 'sandstone' for bedrock, and 'clay', 'till', and 'gravel' for surficial units. High classification performance was achieved with root mean square error (RMSE) scores of 6.42 during cross-validation.

Depth to bedrock spatial prediction was explored using various modeling approaches and feature engineering methods, showing significant effects of spatial features on predictive performance. Spatial lag and spatial neighbors methods outperformed other approaches in terms of reducing spatial autocorrelation and improving predictive accuracy. Feature importance scores revealed that regional variations in relief and physiographic settings, such as valley depth, elevation, and height relative to drainage, were influential across all models.

Spatial neighbors and spatial lags methods were ranked highest in several datasets. In comparing predictive maps, models using spatial lag and spatial neighbors approaches provided the most geologically plausible results. The machine learning approach, particularly the spatial lag feature set, yielded more realistic and detailed predictions than ordinary kriging, especially around major river valleys and local uplands. Finally, comparisons of bedrock topography surfaces derived from machine learning and other spatial interpolation methods demonstrated the superiority of the random forest model in reducing artifacts and capturing the complexity of geological features.

Conclusion

In conclusion, the authors effectively applied a machine learning approach to predict DTB in Alberta, Canada, using terrain morphometry, satellite imagery, and spatial feature engineering. This approach significantly improved prediction accuracy across diverse terrains and produced geologically plausible maps, outperforming traditional methods like inverse-distance weighting and ordinary kriging. The model was fully automated and efficiently handled large datasets without the need for subjective divisions. It offered the potential for further refinement with different spatial weight functions and variations in neighbor usage.