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

• Framework established to examine spatial structure of image registration error.
• Iterative leave-one-out cross-validation method produced local measure of error.
• Spatially variable buffers of detectable change created from interpolated error
• Spatially variable error improves detection of significant lateral migration.
• Broadly applicable for detecting geomorphic changes between features of interest.

Remotely sensed data provides information on river morphology useful for examining channel change at yearly-to-decadal time scales. Although previous studies have emphasized the need to distinguish true geomorphic change from errors associated with image registration, standard metrics for assessing and summarizing these errors, such as the root-mean-square error (RMSE) and 90th percentile of the distribution of ground control point (GCP) error, fail to incorporate the spatial structure of this uncertainty. In this study, we introduce a framework for evaluating whether observations of lateral channel migration along a meandering channel are statistically significant, given the spatial distribution of registration error. An iterative leave-one-out cross-validation approach was used to produce local error metrics for an image time series from Savery Creek, Wyoming, USA, and to evaluate various transformation equations, interpolation methods, and GCP placement strategies. Interpolated error surfaces then were used to create error ellipses representing spatially variable buffers of detectable change. Our results show that, for all five sequential image pairs we examined, spatially distributed estimates of registration error enabled detection of a greater number of statistically significant lateral migration vectors than the spatially uniform RMSE or 90th percentile of GCP error. Conversely, spatially distributed error metrics prevented changes from being mistaken as real in areas of greater registration error. Our results also support the findings of previous studies: second-order polynomial functions on average yield the lowest RMSE, and errors are reduced by placing GCPs on the floodplain rather than on hillslopes. This study highlights the importance of characterizing the spatial distribution of image registration errors in the analysis of channel change.