Re: MemoryFile loses Profile information

Sean Gillies


On Wed, Apr 29, 2020 at 9:40 AM Guillaume Lostis <g.lostis@...> wrote:

Like Ciaran I'm not very familiar with GDAL's python bindings, but I've tried running a little snippet which I believe does the same thing as what was done with rasterio: it writes a JP2 to a vsimem in-memory file, and then writes the bytes to a file on disk.

from osgeo import gdal

mem, out = "/vsimem/B01.jp2", "new_B01.jp2"
gdal.Translate(mem, "B01.jp2")

# Taken from
f = gdal.VSIFOpenL(mem, "rb")
gdal.VSIFSeekL(f, 0, 2)  # seek to end
size = gdal.VSIFTellL(f)
gdal.VSIFSeekL(f, 0, 0)  # seek to beginning
data = gdal.VSIFReadL(1, size, f)

with open(out, "wb") as ff:

When I look at the gdalinfo of new_B01.jp2, it has a CRS and a transform, so I am not able to reproduce the behavior we have with rasterio.

(Side note: Interestingly enough, the pixel values of new_B01.jp2 have slightly changed with respect to the original B01.jp2 file, so there is some data loss somewhere in the process. But maybe that is expected of the JP2 format and could be avoided by passing extra arguments to gdal.Translate? I mainly have experience with writing GeoTIFFs, writing JP2s is new to me)


I've only now remembered (not being a regular jp2 user) that JPEG2000 is a create-copy format (see and as such is not suited for uses cases like

    with"file.jp2", "w", driver="JP2OpenJPEG", ...) as dataset:

Rasterio tries to abstract over the differences between "create" and "create-copy" formats, but might falter in some cases. We're testing JPEG and PNG in the test suite, but not JPEG2000. If you're constructing a dataset from the bands of multiple other datasets, you should probably be using GeoTIFF as a format, it's well suited for this. And then convert to JPEG2000 before uploading to S3 using rasterio.shutil.copy, which calls on GDAL's GDALCreateCopy and is mostly equivalent to gdal_translate.

Sean Gillies

Join to automatically receive all group messages.