The Tweed Mapping Project: machine learning methods for rapid Quaternary mapping

When large areas of the UK were mapped over 100 years ago priority was given to identification of mineral resources. Many such ‘drift’ maps therefore are not consistent with modern scientific understanding, nor do they reflect current stakeholder interests. Surface and groundwater flooding represent a major hazard to homes, infrastructure, and land management across the Tweed catchment. Recent work by BGS Groundwater has indicated that slope deposits are far more widespread than previously identified and play a significant role in groundwater connectivity. Updating the superficial geology map across the ~5000 km² catchment is therefore critical for improving flood forecasting, and the design of a major baseline monitoring project, the Flood-Drought Research Infrastructure funded by NERC. 

The Tweed Mapping Project applies spatial Random Forest models using DTM derivatives at 25 m resolution to predict twelve different deposit classes (e.g. till, alluvium, regolith, talus). Model training data are derived from detailed mapping surveys dated 2005, 2009 and 2012.   

Initial results indicate that slope deposits have been under-mapped, with till being the dominant deposit predicted. Both over and under-sampling are a significant issue; sample adjustment methods are unable to compensate. Minor deposits are therefore under-represented in model outputs. 

Model outputs have been checked in the field in Cheviot, Tweedsmuir and Galashiels areas during 2025. Geomorphological mapping, section logging, and bulk sampling of deposits are being used to provide up-to-date training data to enable more reliable and accurate model predictions. Outstanding issues include: (i) the absence of LiDAR data away from major river channels and settlements, (ii) over-representation of specific field observations, and (iii) limited geomorphological inputs to the model.