Misalignment of landsat images in GRASS?

Question

I am new to Landsat images processing, and I'm trying to create land cover change maps over an area covering around 15 landsat scenes.

I downloaded all scenes available, unfortunately some layers don't align perfectly over time, even having same path and row, and same coordinate system (UTM22N). At first, I just deleted the misaligned scenes, because I had enough images available to get a correct cloud free map for each year I wanted.

But in a few cases (p227r057, p227r058 for example, but from what I've seen until now, I'm facing the problem for probably half of the landsat locations), I cannot find more that 2 or 3 correctly aligned images over more than 20 layers. It is probably just a vertical and horizontal shift, but the problem is that this shift is not constant in distance (from a few tens of meters to a few kilometers) and in direction.

Is there an easy way to cope with that problem, or is it necessary to proceed to the orthorectification of all (a few hundreds scenes multiplied by the number of bands) images?

I'm working with GRASS GIS.

Answer 1

I had a similar problem recently. Older Landsat images are indeed geometrically corrected, but not using GCPs (Ground Control Points), what makes them less accurate. I found an image that was misplaced by even 40 km! This applies to scenes that have L1G or L1Gt processing level. Newer images (like those coming from Landsat 8) that have L1T/L1TP are corrected using GCPs and their position is precise. Unfortunately, there is no way for USGS to correct these data, as there were no GCPs on site in the past. Manual shifting is the only way, but here comes another trap - some are not only shifted, but also warped and they need you to georeference them again (possibly to a L1T/L1TP image in the same path/row). At least bands inside one image have the same displacement.

Answer 2

I am beginning to translate all my layers using shift() function from R package raster. Using a loop it is possible to apply the x and y translation (manually measured) to every bands of a same scene.

So it is still very long but at least it is possible. Alignment gets better but it is still not very accurate.