certain resampling methods fail when CHECK_WITH_INVERT_PROJ is False

[perrygeo]

with rasterio.drivers(CHECK_WITH_INVERT_PROJ=False):
    ...
    # works
    reproject(..., resampling=Resampling.nearest)

    # succeeds but produces blank image
    reproject(..., resampling=Resampling.bilinear)

If CHECK_WITH_INVERT_PROJ=True, both nearest and bilinear work fine.

Related discussion about making invert proj checking the default #613

[sgillies]

@perrygeo this has me rethinking the keyword argument I proposed in #613. Maybe the application of CHECK_WITH_INVERT_PROJ is more pervasive than I realized. Baking True into our environments is making more sense to me.

[perrygeo]

This is a fun one. I parametrized the test to run with all resampling methods and here's the results:

tests/test_warp.py::test_resample_no_invert_proj[Resampling.nearest] PASSED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.bilinear] FAILED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.cubic] FAILED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.cubic_spline] FAILED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.lanczos] FAILED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.average] PASSED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.mode] PASSED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.gauss] FAILED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.max] PASSED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.min] PASSED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.med] PASSED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.q1] PASSED
tests/test_warp.py::test_resample_no_invert_proj[Resampling.q3] PASSED

[perrygeo]

Copied from #753:

Prelim look at gdalwarpkernel.cpp shows that all of the xfailing resampling methods (with the exception of gauss) use their own Filter Functions. Makes sense since the algorithms require a certain radius of neighboring data. But these appear to be failing silently or producing invalid results when CHECK_WITH_INVERT_PROJ is false.

The filter functions themselves are here for future reference.

[perrygeo]

This is a known bug and we can deal with it post-1.0. Getting #780 merged will greatly reduce the chances of users hitting this by accident.

[sgillies]

I wonder if the SOURCE_EXTRA warp option is needed to make some of these resampling methods reliable https://gdal.org/api/gdalwarp_cpp.html#_CPPv4N15GDALWarpOptions16papszWarpOptionsE. I'm going to close this one and we'll reopen if the problems continue with PROJ 6+.