Refining measurements of lateral channel movement from image time series by quantifying spatial variations in registration error

Abstract:

Remotely sensed data can provide information on river morphology useful for examining channel change at yearly-to-decadal time scales. Although the need to distinguish true geomorphic change from errors associated with image registration has been identified, methods such as the root-mean-square error (RMSE) used to assess and summarize these errors fail to incorporate the spatial structure of uncertainty. Here we assess whether observations of lateral channel migration along a meandering channel are significant given the spatial distribution of registration error. An iterative cross-validation approach was used to produce a local measure of error for images in a time series from Savery Creek, Wyoming, USA, and various transformation methods, interpolation methods, and ground control point (GCP) placement were evaluated. Interpolated error surfaces were then used to produce error ellipses that represent spatially variable buffers for the threshold of detectable change. Our results suggest that spatially distributed estimates of registration error allow for detection of more subtle channel changes than using the RMSE or 90th percentile of error in areas where registration is more accurate. Conversely, spatially distributed error prevents changes from mistakenly being regarded as true in areas of greater registration error. Results also support previous findings that second-order polynomial functions on average yield the lowest RMSE and that placing GCPs on the floodplain rather than on hillslopes reduces error. This study highlights the importance of characterizing the spatial distribution of image registration errors in the analysis of channel change.