Continental-scale machine-learning classification of Arctic glacial landscapes using simple morphometrics

Landscapes formed by glacial erosion have previously been classified based on qualitative interpretation of geomorphic evidence, using aerial photography and field studies. In the current era of high-resolution elevation data, limited attempts have been made to improve these classifications using quantitative measurements of landscape form (i.e., morphometry). This is despite landscapes of glacial erosion containing a wealth of information relating to past ice behavior, which can in turn aid our understanding of contemporary and future ice dynamics. This study introduces a new classification method which: (a) has a robust quantitative basis, allowing it to be reproducibly applied to a range of land surfaces; and (b) leverages the power of machine learning to interpret patterns at scale and provide classification probability estimates. The method uses intuitive morphometrics calculated from digital elevation models within a random forest machine-learning model to classify formerly glaciated regions of Arctic Canada and Greenland. The results reveal regional and local variability in glacial erosional style, including distinct populations of scoured landscapes in ice-free parts of Greenland compared to Canada, and wider preservation of non-glacial erosional signatures than previously mapped. Our metrics and classification results show that the degree of glacial modification of the landscape is likely linked to factors including pre-glacial topography, climate (latitude), long-term ice configuration, proximity to ice divides, and geology. Our findings demonstrate the value of simple morphometrics for extracting information from large elevation data sets and provide a quantitative tool for interpreting landscapes whose glacial history is poorly constrained.