Strange behaviour of array_interface on mask and nodata setting

[rhugonnet]

Example below: the nodata has been assigned to 63?!

r.data
Out[8]: 
masked_array(
  data=[[[255, 255, 255, ..., 255, 255, 255],
         [255, 255, 255, ..., 255, 255, 255],
         [255, 255, 255, ..., 255, 255, 255],
         ...,
         [ 74,  76,  79, ..., 121, 119, 141],
         [ 75,  83,  70, ..., 112, 130, 150],
         [ 64,  86,  68, ..., 124, 131, 130]]],
  mask=False,
  fill_value=999999,
  dtype=uint8)
r.data += 5
r.data
Out[10]: 
masked_array(
  data=[[[  4,   4,   4, ...,   4,   4,   4],
         [  4,   4,   4, ...,   4,   4,   4],
         [  4,   4,   4, ...,   4,   4,   4],
         ...,
         [ 79,  81,  84, ..., 126, 124, 146],
         [ 80,  88,  75, ..., 117, 135, 155],
         [ 69,  91,  73, ..., 129, 136, 135]]],
  mask=False,
  fill_value=63,
  dtype=uint8)
np.count_nonzero(r.data==63)
Out[11]: 3441

Looking into it in more details...

[adehecq]

Yes, it's quite common. When not specified, rasterio will set the fill_value attribute of the masked_array as a nodata. See my issue about it here.
The default is I think something like -99999, but when converted to UInt8 (modulo 255), it yields 63.
The best way to overcome this issue is to always set the fill_value and not use the default. But it's a problem for UInt8 when one might not wish to set a nodata (and exclude one value out of 256 only).

[rhugonnet]

In terms of behaviour, I think it should be possible to have a Raster with a masked array that has no invalid data (mask=False) and no fill_value (default 99999). This way we would always mirror the behaviour of np.ma.masked_array.

Here, however, the above behaviour is created by Raster, because np.ma is conserving the original fill_value without modifying it to 63:

test = np.ma.masked_array([0, 255], dtype='uint8')
test
Out[23]: 
masked_array(data=[  0, 255],
             mask=False,
       fill_value=999999,
            dtype=uint8)
test +=5
test
Out[25]: 
masked_array(data=[5, 4],
             mask=False,
       fill_value=999999,
            dtype=uint8)

To solve this, we probably simply need a clause in _overloading_check or data.setter.