Figure 1. a) Overlapping watermasks from two years in red and blue; purple indicates pixels classified as river in both years. b) Rivermasks produced from the watermasks after identifying river pixels and a closure operation. c) Bank aspect (only one image is shown, and the bank pixels are dilated for visibility). d) distance or magnitude of change for all river pixels that eroded or accreted. e) The SWORD centerline overlain on erosion pixels. During vectorization, each pixel is assigned to the closest river centerline node and a summary of geomorphic change is calculated for each centerline node.

We first quantify the uncertainty of our estimates by propagating the water classification errors through our methods. Both the JRC and Pickens datasets include classification error rates. For the JRC dataset, omission and commission errors are estimated for each seasonality class (seasonal or permanent) and sensor (Landsat 5, 7, or 8). Errors are highest in the seasonal omission category, and lowest in the permanent commission for all sensors. There is not as large of a difference between sensors, however the downlink capability during Landsat 5 was limited, and the scan line corrector failure on Landsat 7 limits the quality of the seasonality classifications. The Pickens dataset takes a different approach, quantifying the omission and commission error rates as a function of distance from the land-water boundary for both the Pickens and JRC datasets. We use these distanced-based uncertainties for both datasets because we believe they better represents the sources and patterns of error and uncertainty in classifications. Further, the seasonality classifications in the JRC dataset do not validate well against the high-resolution, manually trained and classified validation data from Pickens et al. (2020).

We apply these error rates to the annual water masks by randomly adding omission and commission errors according to the distance-rate function in Pickens et al., (2020). We repeat this process according to the number of observations in each annual watermask and then average the erroneous watermasks together, simulating the effect of the per image pixel-scale errors to the annual water masks (Figure S1). Each time we analyze two images for change detection, we also create two noisy watermasks and process them with the same river classification and change analysis methods (sections 2.3 & 2.4). The difference between our ‘clean’ watermasks and the ‘noisy’ watermasks is a quantification of the erroneous planform change we would anticipate given no actual change in planform (Figure S2).

Another source of uncertainty stems from variations in river stage. When looking at water surfaces only in planform, changes in inundation cannot be distinguished from changes in channel form. For example, if we happen to observe a flood in year one and low flows in year two, our bank migration data could suggest accretion along both banks. We reduce this source of error by using the composited annual masks, though interannual variability will still be present in our data. We investigate the severity of interannual variation by performing our calculations on 9 combinations of years, in a 3x3 matrix of years from 2000-2002 compared to 2017-2019 (Figure S3). Our final dataset incorporates these data by presenting the minimum, maximum, and median riverbank erosion and accretion values for each node.

To assess the accuracy of our riverbank migration estimates, we calculated the bank erosion rate for river reaches where previous studies have published rates of erosion. After filtering the dataset of 290 published studies (Rowland and Schwenk, 2019) down to studies that obtain similar metrics (producing reach-averaged rates and over decade-scale time intervals) and wider than 150 meters, there are 25 cases remaining (See supplemental S4 for a map of validation sites and Supplemental T1 for full list of references and reach characteristics). We limit the validation studies to decade-scale measurements because year-to-year variations in discharge and erosion can bias or misrepresent the characteristic erosion rate when the time scale of observations changes (Donovan et al., 2019). Further, any change in erosion rate over time could falsely add error to the validation, so we exclude studies with data collection that ended before our satellite record. The reach definition is not identical between the published studies and our data so we compared the migration rate for each validation reach to our observed migration rate nearest the reported validation reach center.