rio-alpha giving a ValueError on an RGBA tif about rio-alpha HOT 2 CLOSED

Also seeing the same bug on an existing 4-band test fixture

$ rio alpha tests/fixtures/ca_chilliwack/13-1326-2805-test-2015-2012_30cm_592_5450.tif /tmp/foo.tif

There's a couple ways to approach this:

As a bug in alpha_mask.mask_exact

We are rolling the axis to the depth (number of bands) when we should be using the number of dimensions which is almost always three. The fact that most inputs happened to have 3 bands was a coincidence that hid this bug. https://github.com/mapbox/rio-alpha/blob/master/rio_alpha/alpha_mask.py#L20-L21

Why np.rollaxis?

If we need to transpose the array to band-last/image axis order, we should use numpy.transpose which is roughly 3x faster and easier to read

np.transpose(img, [1, 2, 0])

Alpha band should take precedent

The bigger question is why are we even doing ndv masking calculations at all if the dataset has an alpha band? Per the conversation at #9 (comment), we should rely on rasterio's heuristics to generate the nodata mask. IOW, the usage of the mask_exact function could be replaced entirely with Rasterio's .dataset_mask() method.

from rio-alpha.

Another related bug, if we fix mask_exact, rio alpha assumes that it can stack the mask on the bottom of the existing array. IOW, If we start with a 4 band array, we get 5 bands which can't be written by rasterio.

from rio-alpha.