Knowledge Bank

Remotely sensed observations have proven to be a vital data source when analyzing river flow, particularly in resource-limited regions. Although remote sensing is an excellent alternative to gathering large amounts of field data, some uncertainty is commonly associated with satellite datasets. River discharge calculated via remotely sensed width, height, and slope is likely biased due to the uncertainty contained in these observations. A primary step in understanding this error is evaluating the uncertainty within these individual variables. Here I took an important step towards analyzing channel width predictions by walking the shorelines of 3 rivers in Alaska: the Knik River, the Tanana River, and the Nenana River. Specifically, I tracked the rivers' channel shorelines using the Bad Elf GNSS Surveyor. Following the post-processing of this data, the tracked shoreline path was compared to a classified water mask of the river, created using 1.2m resolution WorldView satellite imagery. Every 10 meters, segments were drawn perpendicular to the river's centerline, measuring the distance between the tracked shoreline (via Bad Elf) and predicted water pixels using the classified river mask. Overestimates resulted in a positive value, while negative values represented an underestimate of the water mask relative to the Bad Elf path. Overall, the averaged shoreline error, or bias, was an overestimate, at 4.7m. Furthermore, the mean absolute error (MAE) for all rivers was 8.2m, with a standard deviation (SD) of 6.9m. When evaluating the rivers on separate terms, the Knik, Nenana, and Tanana had relative MAEs of 7.0m, 7.6m and 10.3m. Lastly, the percent bias was calculated by dividing shoreline bias by shoreline mean absolute error. The Knik had the highest percent bias of 96.4%, revealing the important role bias played in the shoreline MAE. Shoreline error documented in this study is not identical to width error. If shoreline error is dominated by georegistration error, then width error may be less than shoreline error. If georegistration error is small, then width error may be up to twice as large as shoreline error. These results provide implications for a better understanding of the accuracy of current high-resolution imagery while giving insight into the source of width errors in today's datasets.