Capturing exposed bedrock in the upland regions of Great Britain: A geomorphometric focused random forest approach

Rock exposure distribution maps provide invaluable information for a range of applications from geohazard assessment through to aggregate resource potential assessments. Despite the usefulness of such information, it is only available to a limited extent across Great Britain (GB). Recent developments in the application of machine learning approaches to map exposed rock distribution rely on existing geological and land cover maps as the key input data for model training. We present a catchment-scale approach for delivering high-resolution rock exposure maps for GB mountain terrains. Our application has two objectives: establish a consistent and cross-applicable approach enabling feature identification from elevation datasets; use the results and diagnostics of the application to assist in further environmental process understanding. We utilize manual aerial image interpretation, and a suite of geomorphic terrain variables generated from a 5 m Digital Terrain Model as inputs to a distributed random forest model. Eight separate catchment models were derived from the training datasets using a leave-one-out approach. Aggregated results indicate a model accuracy of 79%, with a relatively high model sensitivity (78%) at the cost of relatively low precision (20%). Variable importance assessment highlighted patterns consistent with expected geomorphic controls on rock exposure related to gravity-driven slope processes in mountain landscapes. These results highlight the potential of multi-variant approaches for high-resolution rock exposure mapping, and lay a foundation for further development, particularly in relation to opportunities for further training data capture to ensure model accuracy. The ability to associate features based on geomorphological variables - indicative of landscape processes including erosion and deposition - presents opportunities that go beyond rock exposure such as for critical mineral and resource assessment. This approach will be applied for initial site characterisation as part of future onshore and offshore geological survey activities where high-resolution terrain and bathymetric data are available. • Bedrock exposure predicted in upland areas of Great Britain based on terrain derivative information with a random classifier model with accuracy of 79% • Possible to estimate rock presence using terrain information alone in the absence of other environmental data, which could provide potential rapid landscape characterization in remote areas, offering a time and cost-effective alternative to field mapping or manual assessment • Dominance of slope, elevation, TWI, MrVBF and standard deviation across the full range of catchment morphologies sampled in predicting rock presence/absence is consistent with gravity-driven slope processes controlling the distribution of soil and other sediment deposits on hillslopes • The approach presented here can readily be adapted to other terrain related classification problems such as in the fields of geochemical mapping, habitat modelling, and superficial deposit mapping amongst many other opportunities • Underlying random forest classifier code base is available via Github ( https://github.com/BritishGeologicalSurvey/geospatial-rf/releases/tag/v1.0.0-external ) under GNU Lesser General Public License v3.0.