On Wed, Jul 4, 2018 at 4:09 AM, Denis Rykov <rykovd@...> wrote:

Hi there!

I have a general question about reprojection. Consider the following example:

with rasterio.open(fname_src) as src:

    with rasterio.open(fname_dst, 'w', **kwargs) as dst:

        for i in range(1, src.count + 1):

            rasterio.warp.reproject(

                source=rasterio.band(src, i),

                destination=rasterio.band(dst, i),

                src_transform=src.transform,

                src_crs=src.crs,

                dst_transform=dest_affine,

                dst_crs=dst_crs,

                resampling=resampling)

As I can see from source code rasterio reads each band as an N-d array.
Does it mean that rasterio loads each band into memory?

My main question: can I use this code for warping rasters of arbitrary size (or I'm limited by memory size)?

Thanks.

The source dataset's bands are incrementally loaded into memory as the warper works, but only up to the limit of GDAL's block cache (Rasterio uses the GDAL C library). If you profile your program and see the memory allocation increase, this is due to GDAL's caching. You can see also the notes at http://trac.osgeo.org/gdal/wiki/UserDocs/GdalWarp#WarpandCacheMemory:TechnicalDetails, which apply to Rasterio.

Please note that you can warp multiple imagery bands in one reproject() call by passing sequences of band indexes.

  reproject(source=rasterio.band(src, src.indexes), destination=rasterio.band(src, dst.indexes), ...)

This is faster than looping over the bands.