Save geotiff with keep_palette still add an alpha band

[steph-ben]

Describe the bug

When saving an image with the code below, I expect only one band to be written on my Tiff file.
However, at one point my XRImage has mode = "L", and thus not respecting my "keep_palette" argument ..

Any thoughs on this ?

Might be related to #640

To Reproduce

            crop_channel.save_dataset(channel,
                                      writer='geotiff', filename=str(fp), 
                                      # Add our metadata
                                      tags=tags,
                                      # Ensure data type
                                      dtype=np.uint8, floating_point=False,
                                      # Don't transform our data
                                      enhance=False, keep_palette=True,
                                      # Write compressed
                                      num_threads=8, COMPRESS="LZW")
            crop_channel = None

My output looks like :

$ gdalinfo tmp/plop/himawari_B01_20220708_090000.tiff 
Driver: GTiff/GeoTIFF
Files: tmp/plop/himawari_B01_20220708_090000.tiff
Size is 2800, 1600
...
  name=B01
  platform_name=Himawari-8
  sensor=ahi
...
Band 1 Block=256x256 Type=Byte, ColorInterp=Gray
  Mask Flags: PER_DATASET ALPHA 
Band 2 Block=256x256 Type=Byte, ColorInterp=Alpha   <<<<<<------------ I don't want this

By following step by step, I got to the point that :

• My XRImage is created with mode = "L" by default, somere around here : https://github.com/pytroll/satpy/blob/main/satpy/writers/__init__.py#L443

• At XRImage.finalise() https://github.com/pytroll/trollimage/blob/main/trollimage/xrimage.py#L1110

keep_palette = True
self.mode = "L"

and therefore still adding an extra band

Environment Info:

• OS : WSL Debian
• Python 3.9.13
• satpy=0.36.0

[djhoese]

Side note: I see you have floating_point=False, I think that flag has been deprecated for a while and could be removed.

What composite/product are you saving? I see AHI B01 in the geotiff metadata. For keep_palette to work, you need to enhance your data with the palettize operation. This applies a colormap to your data, but instead of converting it to RGB/A your data becomes the index in the colormap for the assigned color. Unless you are doing this yourself outside of save_dataset then there is no palette for trollimage/satpy to "keep". The if statement you pointed to in xrimage.py is doing this check by saying "Does the user want to save the palette/colormap and we have a P(alette) mode image?".

This is sort of described here:

https://github.com/pytroll/trollimage/blob/00c14a660d1be4d21c5be413c9b7d799a0ba5d0f/trollimage/xrimage.py#L1101-L1107

If you are not worried about colormaps and you just want an L image with no Alpha band then you could specify instead fill_value=0 which would keep your image as L but replace all NaN or invalid pixels with 0 (black).

This is described in multiple places for better or worse, but probably in most detail here:

https://github.com/pytroll/trollimage/blob/00c14a660d1be4d21c5be413c9b7d799a0ba5d0f/trollimage/xrimage.py#L1085-L1097

[steph-ben]

Thanks a lot, I think fill_value=0 was exactly what I missed !

Actually in my processing I do something like that :

scn = Scene(reader='ahi_hsd', filenames=files)
...
ds = scn[channel]
dataset_scaled = xr.apply_ufunc(scaling_func, ds)
...
crop_channel.save_dataset(channel, writer='geotiff', filename=str(fp), fill_value=0, ...)

So yes I'm already converting my data from albedo values to pixel, following an internal mapping.
My output needed to follow some specific Tiff convention, and the two bands were causing errors.

PS : Many thanks for the quick answer and for this amazing lib ! Automatic generation of pre-defined composite is really a killer

[djhoese]

Glad that worked! Let us know if you have any problems and thanks for the kind words.