Climate interacts with the Earth’s surface in several ways: It directly shapes the surface by river incision and hill slope processes (dependent on precipitation amount and characteristics), frost cracking (dependent on temperature and water availability) and aeolian (wind induced) erosion; it indirectly shapes the Earth’s surface by modifying vegetation cover and type, which in turn affects surface erosion. Detecting patterns in erosion-relevant climate change through geological time helps provide climatic context for erosion rates calculated from geological archives; machine learning techniques applied to palaeoclimate simulations can help identify and explain these patterns (Fig. 1).
Fig. 1 The multivariate anomaly maps for pre-industrial (PI) and Last Glacial Maximum (LGM) comparisons show the geographical coverage of clusters C1-C6 in Western South America, which describe the spatial extend of regions characterised by similar modes of change. The corresponding modes of change (b) for each cluster are expressed as relative changes in each of the 9 investigated variables: 2m air temperature (te2m), 2m air temperature amplitude (t2am), consecutive freezing days (csfd), freeze-thaw days (fthd), maximum precipitation (pmax), consecutive wet days (cswd), consecutive dry days (csdd), zonal near surface wind speeds (u10) and meridional near surface wind speeds (v10). The score (b) expresses the goodness of discriminability between the palaeoclimate pair PI-LGM in each of the anomaly clusters. The size of the circles corresponds to the relative contribution of each of the 9 climatic attribute variables to the measured discriminability in each anomaly cluster for all three time slice comparisons. This figure is published in Mutz and Ehlers (2019).