Apply same resampling method for mask and tile

[vincentsarago]

in

rio-tiler/rio_tiler/utils.py

Lines 345 to 353 in 2bf9f14

	with WarpedVRT(src_dst, **vrt_params) as vrt:
	data = vrt.read(
	out_shape=out_shape,
	indexes=indexes,
	window=out_window,
	resampling=Resampling[resampling_method],
	)
	mask = vrt.dataset_mask(out_shape=(tilesize, tilesize), window=out_window)

we could end up (actual behavior) with using different resampling for mask and data which will result in bad masking.

[vincentsarago]

@dionhaefner
Am I missing something here ? it seems to be the same in terracotta https://github.com/DHI-GRAS/terracotta/blob/master/terracotta/drivers/raster_base.py#L524-L529

[vincentsarago]

Confirmed!

[vincentsarago]

closed by #106

[dionhaefner]

IIRC this was deliberate. We used to use the same resampling for the alpha mask, but it eroded the tile edges. E.g. linear resampling doesn't make much sense on a binary mask, does it?

[vincentsarago]

@dionhaefner well I'm not sure how it does on a pure binary mask but at least when dealing with mask built from nodata it will mater.
Will need to test on pure alpha/binary mask

[dionhaefner]

When I played around with this, using resampling for the mask never had a positive effect. But YMMV

[vincentsarago]

Well it seems that this introduced a bug when dealing with internal mask (wonder why the tests didn't caught this)

When using bilinear resampling the mask is =255 everywhere

[vincentsarago]

Well this is really strange I could not reproduce the bug (bilinear resampling of mask creating non-masked data) when using Rasterio wheels ... so I guess this has to deal with my gdal/numpy install 🤔 .

BUT I found that using something else than nearest resampling introduce mask artefacts ... I'm going to revert the change introduced in #106

I'm really sorry for the users that could have been impacted by the 1.2.6 release

[vincentsarago]

Note: if we revert #106
this test :

rio-tiler/tests/test_utils.py

Lines 753 to 756 in 156b0a8

	tilesize = 128
	arr, mask = utils.tile_read(address, bounds, tilesize, nodata=0)
	masknodata = (arr[0] != 0).astype(np.uint8) * 255
	np.testing.assert_array_equal(mask, masknodata)

fails which seems a bit strange. I'll investigate a bit more and might open an issue on rasterio

[vincentsarago]

closed by d9e532a