Riverbank migration has historically been seen as a risk to infrastructure that can be combatted through channelization, bank stabilization, and sediment trapping. The physical processes involved with riverbank erosion and deposition are well defined, yet the solutions to these equations are computationally and data intensive over large domains. While current understanding of large-scale river channel mobility largely comes from reach- and watershed-scale observations, we need global observations of riverbank erosion and accretion to understand sediment processes within and across river basins. In this work, we create the first global dataset of riverbank erosion for >370,000 kilometers of large rivers using 20 years of water classifications from Landsat imagery. We estimate uncertainty by propagating water classification errors through our methods. Globally, we find riverbank erosion for rivers wider than 150 m to have an approximately log-normal distribution with a median value of 1.52 m/yr. Comparing our dataset to 25 similar estimates of riverbank migration, we found an normalized mean absolute error of 42% but a bias of only 5.8%. We definitively show that river size is the best first-order predictor of riverbank erosion, in agreement with existing literature that used available data. We also show that the relationship between size and bank erosion is substantially different among a sample of global river basins and suggest that this is due to second-order influences of geology, hydrology, and human influence. These data will help improve models of sediment transport, support models of bank erosion, and improve our understanding of human modification of rivers.