The results of our classification uncertainty analysis show that the global median bank migration rate (1.52 m/year) is more than an order of magnitude higher than the median false erosion rate associated with pixel misclassification error (.06 m/year). The distribution of the observed riverbank migration rate is approximately log normal, while false erosion rate is slightly bimodal when binned in log scale (Figure 2b). Of course, it is not true that the bank migration rate is an order of magnitude higher than the false erosion rate everywhere, in fact many reaches with low migration rates will have nearly equivalent rates of false migration. 5.7% of reaches have false erosion rates higher than their median observed bank erosion rate. Narrow rivers and anabranching rivers have more bank pixels per unit length than wide single channel rivers, and as such will have higher average rates of classifcation errors. The uneven distribution of Landsat images also greatly biases the false erosion estimates, as areas with fewer images have higher rates of classifcation errors in the annual watermask products (Figure S3).
Figure 2 . a) Validation of the riverbank erosion from REAL (vertical axis) with references from literature (horizontal axis). b) The distribution of global observed riverbank erosion rate is shown in blue. The false erosion rate is shown in orange and indicates the uncertainty from water classification errors.
Our global estimates of riverbank migration are mapped in Figure 3a after averaging the data to Pfaffstetter level 6 basins. At the largest scales, these data agree with our expectations: rivers like the Amazon, Orinoco, ParanĂ¡, Mississippi, Syr Darya, Indus, Ganges-Brahmaptura-Meghna (GBM), Yangtze, and Niger show very high rates of erosion. In contrast, rivers in Europe, much of Siberia, and the Canadian Shield show low average rates of bank migration. European rivers have been modified and stabilized since at least the 16th century to serve the people that live near them, and as such show very low rates of riverbank migration (Evette et al., 2009; Grizzetti et al., 2017). The Canadian Shield region also shows near-zero riverbank migration rates, but for very different reasons. Rivers in the Candian Shield region are not heavily modified by humans, but flow over highly resistive precambrian igneous and metamorphic rock. Large swaths of the Arctic show low migration rates, with some notable exceptions in the Russian Far East, Northern Siberia, and parts of the Yukon River.
At finer scales we can see intrabasin and reach scale patterns. Within the Amazon basin, the subbasins coming off of the Andes mountains show the highest rates of erosion, while tributaries from the central trough are more stable (Figure 3b). Upon closer inspection, and aggregated to the 10km SWORD reaches, we can see high variability in bank erosion between reaches, with a distinct drop in erosion downstream of a confluence with a tributary (Figure 3c). Visual inspection of satellite images suggests that this tributary (Ituxi River) typically has lower sediment concentrations than the Purus River. At the finest scale of REAL, the 200m node spacing, we see meander-scale sediment processes on the Purus River (Figure 3d). Nodes near the apex of the meander bends typically show high rates of erosion, while the straight connecting sections show much lower rates of erosion. We also see a pattern of high rates of erosion before and after the apex of the bend, indicating the expected positive bias in erosion on the downstream bank faces.