Multiband nodata masks

[BielStela]

Hi!
I found a problem with how the nodata mask is handled for multiband rasters. It happens that I have a multiband raster with no internal mask, only a global nodata value, and I expect that when asking for a tile of a specific band, the mask is the one for that band. Instead what I get is the mask for the first band only. Looking at the code I see that deep in the low level read function uses

data = dataset.read(
    indexes=indexes,
    window=window,
    out_shape=(len(indexes), height, width) if height and width else None,
    resampling=resampling,
    boundless=boundless,
)
mask = dataset.dataset_mask(
    window=window,
    out_shape=(height, width) if height and width else None,
    resampling=resampling,
    boundless=boundless,
)

Is there any major reason on why multiband rasters can have only one mask representation for all bands?

I'm sure it is not as simple as to use dataset.read_masks with the indexes param because it will return a mask with an extra axis and that will break lots of things down the line. But maybe is it something worth investigating.

Thank you!

[vincentsarago]

@BielStela see https://github.com/rasterio/rasterio/blob/main/rasterio/_io.pyx#L991-L1010 for how the mask is created from multi band dataset with internal nodata value.

Instead what I get is the mask for the first band only.

are you sure, to me it seem the mask should be the combination (https://github.com/rasterio/rasterio/blob/main/rasterio/_io.pyx#L1035) of all band's mask 🤷‍♂️