REMOTE PREDICTIVE MAPPING OF SURFICIAL GEOLOGY USING RANDOM FOREST CLASSIFICATION ALGORITHMS: EXAMPLES FROM NORTHERN CANADA

Remote Predictive Mapping (RPM) has been used by the Geological Survey of Canada as a mapping strategy capable of quickly mapping large areas at relatively low cost. Although its use is not widespread, RPM efforts have yielded encouraging results in terms of the applicability of the method and the reliability of the mapping results. This presentation summarizes some of these mapping efforts and the methodological developments achieved during mapping of surficial materials and landforms for NTS Mapsheet 75M (Mackay Lake).

The RPM map results from classification of satellite imagery (Landsat ETM+), using training areas to associate distinctive spectral signatures with unique material types. Initial image classification relies on identification of moisture content differences within the landscape. Moisture content is used as a proxy measurement incorporating multiple terrain characteristics such as topographic position, sediment thickness, and grain size. Moisture content measurements are converted to geologic materials based on a series of rules-based decisions incorporating spectral texture, understanding of glacial landform genesis, landform associations, and topographic position of classified pixels. The final map is a pixel-based map (30 m pixels) depicting the distribution of bedrock, bedrock-rich areas with surface boulders, and the gradational character of the thickness of glacial sediments on the landscape.

Image classification was performed using the Random Forest statistical algorithm complemented by additional stochastic training and validation steps. Classification refinements were completed using a heuristic expert system. Classification outputs provide an estimate of overall accuracy and also allow for additional derivative products such as 'reliability maps' depicting the the spatial variability of classification accuracy for each variable. These classification derivatives provide insights into the robustness of training data, and classification performance. In turn, they help target areas requiring more detailed mapping and field checking.