Mapping surface deposits with machine learning

In Sweden, detailed Quaternary deposit maps cover only about one-third of the country. This thesis examined whether machine learning and deep learning can accelerate surface deposit mapping and improve depth-to-bedrock prediction, while evaluating the opportunities and limitations of a data-driven approach. Surface deposit classification was tested using Extreme Gradient Boosting and multi-view evidential deep learning, while a two-part framework was developed to separate bedrock outcrop classification from continuous depth prediction, providing spatially explicit uncertainty estimates. In addition, historical land use was classified from scanned historical maps. Performance varied considerably by deposit type. Peat and bedrock outcrops were reliably identified and are promising candidates for automation. Till achieved high aggregate performance, but the most confident till predictions were frequently incorrect. Sorted sediments remained beyond the reach of current approaches. Depth-to-bedrock predictions were accurate within 10 m, though the model was less precise in this range. At greater depths, the model increasingly underestimated depth while becoming overconfident in its predictions. A recurring finding was that machine learning and existing map products are best viewed as complementary. Machine learning offers higher spatial resolution and explicit uncertainty estimates, but struggles with deposit classes and depth ranges where the link between surface data and the target is weak. It is therefore better positioned as a tool that supports, rather than replaces, expert mapping. While producing uncertainty estimates is technically feasible, bridging the gap between their production and practical use remains an open challenge.