I am working on processing stacks of S1 GRD images and have run into some issues with georeferencing. We are processing each scene over a subsetted region of interest with an identical workflow using an external DEM (Read --> Thermal Noise Correction --> Calibration --> Terrain Flattening --> Terrain Correction --> Subsetting). We are finding that the resulting images do not align perfectly despite the fact that we are reprojecting them into the same coordinate system. Is there a way to force the georeferencing to project the images onto an identical grid without subpixel-pixel offsets? We are trying to reduce the number of steps needed to process and to avoid any unnecessary image resampling. Is the only way to do this in SNAP by using the collocation/stack tools? Listed below are a few example outputs that show the small differences for corner reference points and pixel spacing for images over the same area.
map info = {Geographic Lat/Lon,2790.0,2790.0,51.50009338883949,36.49986663855602,1.7966305682390457E-4,1.7966305682390457E-4,WGS84,units=Degrees}
map info = {Geographic Lat/Lon,2790.0,2790.0,51.50000894422885,36.49993037646338,1.7966305682390457E-4,1.7966305682390457E-4,WGS84,units=Degrees}
map info = {Geographic Lat/Lon,2790.0,2790.0,51.500117638116166,36.499833068046414,1.7966305682390457E-4,1.7966305682390457E-4,WGS84,units=Degrees}
map info = {Geographic Lat/Lon,2790.0,2790.0,51.500176926885864,36.49986741898079,1.7966305682390457E-4,1.7966305682390457E-4,WGS84,units=Degrees}
As SAR images are acquired from slightly different incidence angles (or in case of ascending and descending passes even different directions) smaller shifts in image geometries are quite common. You can solve this by coregistering your products, either after terrain flattening (and then terrrain correct the coregistered stack) or by directly coregistering the terrain corrected subsets.
You have to define one master image to which all other products are aligned.
Arbitrary pixel grids are not useful and always require additional resampling . Why incur additional processing time, and slight further degradation of the image by additional resampling?
Many users work with data from different sources, and yes if the pixel sizes are not multiples then resampling is required. However, many work with 10 m or 20 m S2 and 5 m or 20 m S1 data that is coregistered when a uniform global pixel grid (although different size) is used. With a single pixel size, all users’ products will precisely overlay, whereever produced using SNAP.
We implemented global grids for all our resampling 20 years ago. The result is that everything produced by all users any time in the last 20 years precisely aligns for a specific pixel size. There are no subpixel shifts.
Global grids are defined as integer multiple of the pixel size from the origin of the projection or equator and prime meridian.