I acquired several Landsat 5-8 images corrected at Level 1T. I also have aerial ortophotos of the same area. From a visual assessment the co-registration between the images seems good, but I need to quantify the registration accuracy. My approach is to simply find some Check Points and calculate the average RMSE, but which transformation type should I choose to compute the residuals? In the Landsat product specifications and user handbooks I can't seem to find specific info about what they use. (I am using QGis 2.10)
The standard for computing the registration accuracy is to consider the euclidian distance between homologous check points to estimate your RMSE (they are the residuals of the underlying model used to geometrically correct your image).
Those check points should be selected on location that are stable in time (e.g. avoid the boundaries of temporary water bodies) and (except for true ortho from high precision DSM) where there is no abrupt elevation difference with the neighbours.
As mentionned in the comments, the final RMSE is independent from the underlying model used to create the orthorectified image.