Can't join multiple DEMs using QGIS 'Build Virtual Raster' if they span 2 CRS's

Question

I have 20 odd DEMs downloaded from https://search.asf.alaska.edu/ (ALOS PALSAR)

Location is the Falkland Islands.

Here is an example just adding 2 of the tiles into QGIS which differ in terms of their CRS

EPSG:32720 - WGS 84 / UTM zone 20S

EPSG:32721 - WGS 84 / UTM zone 21S

As soon as I try putting them into a Virtual Raster using following settings:

I get only a single DEM (20S CRS) within the Virtual Raster which would probably make sense:

Switching Build Virtual raster settings to 21S for 'override projection'

And the same 20S raster just shifts to the right, no sign of 21S tile:

If I turn on 'allow projection difference:

I get the 21S tile flipping the other side of the 20S tile

Is it just not possible to have DEM tiles that cross 2 CRS's within a Virtual Raster, and if not is there another way?

Answer 1

Thanks B-C B I figured out a method to resolve:

Adding test selection of DEM's from the following ALOS PALSAR dataset into QGIS:

As you can see I have only 1 DEM tile on 20S CRS, so I'll see if I can simply re-project that one to match the others with 21S

First switch QGIS project CRS to 21S (bottom right globe icon)

Create shapefile for square extent in this case 250000 x 250000

Extent: (229085.941613, 4146747.332439) - (479085.941613, 4396747.332439)

Run the Warp process on the single 20S DEM tile

Layers over perfectly:

Now run Build virtual Raster selecting only the assets that are EPSG:32721

For some reason QGIS can't render this new vrt very well:

But we will push forward and use the gdalwarp command to cut the new virtual raster above to the square extent created earlier:

Not forgetting to switch the CRS to 32721 which is what the virtual raster is set to:

gdalwarp -t_srs EPSG:32721 -wo SOURCE_EXTRA=1000 -tr 12.5 12.5 -srcnodata "-9999" -r cubic -ot Float32 -of GTiff -te 229085.941613 4146747.332439 479085.941613 4396747.332439 "F:/GIS/Falklands/virtual_final.vrt" F:/GIS/Falklands/dem_final.tif

All working nicely on UTM 21S CRS:

I'm not too happy having to process a DEM up to three times to get to a clean unified extract of the full dataset, so I'll keep looking for an alternative, potentially doing it all via GDAL commands. But degradation of the data seems to be kept to a minimum, likley helped by using cubic convolution resampling.