Python GDAL – How to Crop Raster in Memory Using GDAL Bindings

gdalgdalwarppython

I have a bunch of rasters I need to crop to a bounding box, perform come calculations with, and then close; I have no need to keep the cropped raster.

So far I've only been able to crop using gdal.Warp(dest_file, src_object, bbox), which creates the unwanted dest_file.

I'd like dest_file to point to an empty gdal object, but can not figure out how to initialize one.
I've created one in memory using:

mem_drv = gdal.GetDriverByName('MEM')
dest = mem_drv.Create('', rast.RasterXSize, rast.RasterYSize, 1, gdal.GDT_Float32) 
dest.SetProjection(rast.GetProjection())
dest.SetGeoTransform(rast.GetTGeoTransform())

where rast is an raster opened with gdal.Open

This lets me store the output of gdal.Warp in memory but it does not crop it; I'm assuming because the width is hard coded in the creation routine.

What is the best way to store the cropped image in memory?

Should I just get the pixel spacing from the original raster and initialize the size of dest using this?

Best Answer

With GDAL 2.2.0+ use the VSIMEM filesystem (http://www.gdal.org/gdal_virtual_file_systems.html). It allows you to treat blocks of memory as a file within the virtual file system. This version uses the projWin parameter of gdal_translate to clip from a bounding box.

gdal.Translate('/vsimem/clip.tif', 'path/to/input.tif', projWin=[ulx, uly, lrx, lry])

You can then open the output using the standard approach:

ds = gdal.Open('/vsimem/clip.tif')

Note that Translate was used over Warp because Warp can get buggy when specifying a gdal dataset object as the input. Warp is also more traditionally used for clipping along vector features, Translate is simple the better option for clipping from a bounding box.

Using the VSIMEM filesystem is a simpler option than creating your own memory dataset to store the output because it is directly taking the output of Translate and writing it into memory. On the other hand, any discrepancies between your output clip and your instantiated memory dataset will result in weird interactions.

Related Question