Continental-scale machine-learning classification of glacial landscapes using simple morphometric parameters

Around 50 years ago, David Sugden and Brian John proposed a classification scheme for landscapes formed by glacial erosion, based on the geomorphic evidence from the beds and peripheries of Quaternary ice sheets. This evidence was interpreted qualitatively using aerial photography and limited field visits. In the current era of ubiquitous, freely available, high-resolution elevation data, limited attempt has been made to update this classification using quantitative measurements of landscape form (i.e., morphometry), and no such scheme has been applied to the Northern Hemisphere regions where the scheme was originally developed. This is despite landscapes of glacial erosion containing a wealth of information relating to past ice behaviour. This study therefore aims to create a new classification method which: (1) has a robust quantitative basis allowing it to be reproducibly applied to new land surfaces, including those currently buried beneath modern ice sheets; and (2) takes advantage of the power of machine learning approaches to interpret patterns at scale and provide estimates of classification confidence. The method presented here uses intuitive morphometrics that can be calculated relatively simply from digital elevation models, including total relief, spatial density of local peaks and basins, and drainage characteristics. A random forest classifier is trained on a selection of manually classified landscapes and the resulting model is used to reclassify the exposed regions of northern North America. Classification confidence is examined using decision tree voting scores, and ‘out of bag’ error values are used to estimate variable importance. The results align broadly with the original 1970s classification scheme, but reveal more local- to regional-scale and intraclass variability than was previously accounted for. Low confidence scores allow identification of landscapes which represent a more complex interplay of different erosional styles, or a transition between classes. The classification also highlights the preservation of a range of non-glacial erosional signatures, even in areas known to have been affected by Quaternary glaciations. Overall, the work demonstrates the value of simple morphometric parameters for condensing information from large elevation datasets and provides a quantitative tool for interpreting landscapes whose glacial history is poorly constrained, such as those hidden beneath modern ice sheets, or on other planets.