Padding to original bounds extends the array

[davidbrochart]

Code Sample, a copy-pastable example if possible

import rioxarray
import xarray

xds = xarray.open_dataarray("test/test_data/input/MODIS_ARRAY.nc")
# xds.shape is (200, 200)
bounds = xds.rio.bounds()
xdsc = xds.rio.pad_box(
    minx=bounds[0],
    miny=bounds[1],
    maxx=bounds[2],
    maxy=bounds[3],
)
# xdsc.shape is (201, 200)

Problem description

I think we shouldn't include the half-pixel here.

Expected Output

The shape of the padded array should remain the same.

[snowman2]

Does pad_xy behave as expected?

[davidbrochart]

Yes, but I guess it's a question of definition of the bounding box. In pad_box, do we expect a box that is relative to the "pixel boundaries" or the pixel centers (which correspond to the array coordinates)?

[snowman2]

do we expect a box that is relative to the "pixel boundaries"

I would assume that would be the behavior. What are your thoughts?

[davidbrochart]

I don't know, but it has to be consistent with clip_box. It looks like clip_box expects a box relative to the "pixel centers" (it extends it with a half-pixel to the "pixel boundaries"), so in this case pad_box behaves similarly (although clip_box uses rasterio.windows.from_bounds).
Calling pad_box with the rio.bounds() (which include the half-pixel) can lead to a random behavior because of floating point approximations, so it's not a good idea. It's better in that case to call it with the first/last coordinates.

[snowman2]

One idea I had was to reduce the buffer by multiplying the cell size by 0.9999 and see if that helps with the floating point precision issues.

[davidbrochart]

I'm not sure, it would be kind of a hack. I think the current behavior is OK, we might just have to document it a bit more.

[snowman2]

That sounds good to me 👍