The bedrock topography of Alberta is the unconformity surface at the top of Upper Cretaceous and Paleogene bedrock, and reveals geomorphic features created by Paleogene to Recent river systems and Quaternary glaciation.
This grid is a digital rendering of the bedrock topography of Alberta, produced using machine learning techniques to include 39 terrain and geological input predictors to supplement point data. Previously published information from Alberta Geological Survey (AGS) maps and reports, as well as new data were used as input point data. The grid was modelled using a 500 m grid cell spacing and should not be used for local-scale studies requiring a higher spatial resolution.
This grid updates the topographic surface portrayed in AGS Map 602, Bedrock Topography of Alberta. The bedrock topography gridded surface provides the lowermost bounding surface for modelled Neogene-Quaternary sediment units, and is the uppermost bounding surface for bedrock units modelled in the Geological Framework of Alberta. Improving the accuracy of this surface is necessary, as knowledge of the position and thickness of overlying sediments is critical for management of groundwater, aggregate, and energy resources, as well as for other land-use applications.
Step 1: (Input point data) A bedrock topography pick is a point defined in three dimensions (X, Y, and Z). Picks were compiled from previous AGS projects and newly generated data created by AGS geologists. The compiled bedrock topography picks were defined from drillholes and site descriptions collected by AGS geologists, geophysical logs from oil and gas wells, historical information from AGS reports, university theses as well as reports written by external geoscience professionals, geotechnical boreholes submitted by industry partners, outcrops located on provincial imagery, water well records from the Alberta Water Well Information Database, contour information compiled from existing bedrock topography and hydrogeological maps of varying quality, and synthetic control points. Points that were considered outliers by an AGS geologist were removed from the dataset. The remaining 141,882 bedrock topography picks were used as input into the machine learning algorithm (outlined below).
Step 2: (Input predictors) A stack of raster-based predictors were used to provide auxiliary information to support the bedrock topography mapping by using statistical relationships between the geological and geomorphological characteristics of the region and sediment thickness. A 1:20 000 scale provincial DEM was used as the primary source of topographic information. This DEM was used to generate a suite of morphometric derivatives based on scripts using the ‘Rsagacmd’ package in the R statistical programming language, used in conjunction with the open-source SAGA-GIS software. The calculated morphometric parameters include those that describe: (1) local morphology; (2) regional relief characteristics; and (3) topographic wetness. The terrain-related predictors were augmented with predictors that allow spatial autocorrelation to be accounted for in the modelling process, including x/y spatial coordinates and known buried channel thalwegs.
Step 3: (Machine learning) The Extremely Randomized Trees (XRT) algorithm in the R package ‘ranger’ (Wright and Ziegler, 2017) was used for the spatial prediction. XRT is a type of Random Forest model that constructs a ‘forest’ of uncorrelated decision trees on random subsets of the training data. Each tree produces a prediction and the final prediction is chosen by averaging (for regression) or majority voting (for classification). In contrast to traditional Random Forests, XRT further attempts to reduce model variance by choosing from several randomly drawn thresholds in each predictor when splitting a node. The XRT model was trained from the pick data using the measured thickness of the unconsolidated sediments as the response variable and the values of the raster-based predictors at the same locations. These predictors were augmented by additional distance-based features and spatial lag variables from neighbouring pick locations. A forest of 500 trees was used in the model, and the hyperparameters specifying the number of predictors that are randomly available for each split, and the minimum number of observations that are permissible in a leaf-node were optimized using nested cross-validation with a grid-search strategy.
Step 4: (Bedrock topography) The bedrock topography was derived by subtracting the modelled sediment thickness (using step 2 and 3 above) from the land surface digital elevation model. The land surface DEM was generated at a 500 m grid cell size using the aggregate tool in ArcGIS to resample the DEM at a coarser resolution that would be appropriate for a provincial-scale model. The bedrock topography gridded surface was rendered at the same 500 m grid cell size resolution, which is intended for provinical-scale investigations and not applicable for site specific studies. The bedrock topography surface grid was produced in Transverse Mercator projection using the North American Datum 1983 (NAD83).