Unexpected behavior: st_transform a raster and transforming back, it is not identical unless applying st_warp about stars HOT 7 CLOSED

stars still considers the object to be in a curvilinear grid (which doesn't make sense imo because it isn't anymore)

yes, because it does not contain information about what the original CRS was, in which it was regular

it's plotted slightly different on mapview (surprising)

mapview will do some warping, which is lossy

from stars.

yes, because it does not contain information about what the original CRS was, in which it was regular

hmm, doesn't it? I thought I've passed it this information.

# transforming back
back = transf %>% st_transform(st_crs(temperature))

and, this is TRUE:

identical(st_crs(temperature),st_crs(back))

maybe I misunderstand what you mean here.

mapview will do some warping, which is lossy

I would expect identical warping results of identical matrix data and CRS, which I indeed get, but just after using st_warp.

from stars.

In transf %>% st_transform(st_crs(temperature)), trans doesn't know its grid can be represented by a regular grid in st_crs(temperature).

from stars.

So that means, transforming a raster to another CRS always comes with some sort of informational loss regarding the grid, which cannot be recovered without warping the data? or is there a way I can provide that information when re-transforming?
I have difficulties understanding the reasoning behind it.

At least, it seems somewhat inefficient that the second line in the example below takes overproportionally long to compute, given that in the end, the matrix data is exactly identical. I admit it may seem like an unrealistic special case to do that..

back = transf %>% st_transform(st_crs(temperature))
back_w = transf %>% st_transform(st_crs(temperature)) %>% st_warp(dest = temperature)

So I would suggest the following behavior: when (re-)transforming to the CRS that a raster had originally, assume a regular grid again, as no re-calculation / warping is required in fact. Possibly store the original CRS (I assume it would be stored somehow already).
Or alternatively, add a special case to st_warp that skips the re-computation of the matrix in that case.

from stars.

So I would suggest the following behavior: when (re-)transforming to the CRS that a raster had originally, assume a regular grid again, as no re-calculation / warping is required in fact. Possibly store the original CRS (I assume it would be stored somehow already).

Yes, of course one can do all that, the question is how large the benefit is compared to the cost, and how this can be made waterproof such that all kind of operations in between (cropping, downsampling) lead to the right regular grid again. In your case, going from CRS A to B to A and no messing, you can set back the dimension table by

st_dimensions(back) = st_dimensions(temperature)

but only because you know that this is OK: no cost, also no guarantee or checking from the software side.

from stars.

Ah, that was the missing piece of the puzzle, thank you very much, that solves my problem.

and how this can be made waterproof such that all kind of operations in between (cropping, downsampling) lead to the right regular grid again

I agree, that may be tricky. Thanks for explaining!

from stars.

(can be closed)

from stars.