rioxarray.open_rasterio and xr.open_dataset have different defaults for the mask_and_scale kwarg

[etienneschalk]

This can be confusing when using one or the other. Is there a particular reason for the different default?

Also, even when a DataArray does not have add_offset nor scale_factor attributes set in their encoding (removed explicitly from the attrs), it seems an automatic scale_factor of 1.0 and add_offset of 0.0 are generated (not sure if it is upon writing nor reading?) that, if applied, will likely convert the data to floats, which is unwanted, for instance if working with uint8, I do not want it to be converted to float32. So even if mask_and_scale where enabled, the absence of such attributes should make the mask_and_scale a noop and preserve the dtype.

Note: I actually prefers the rioxarray.open_rasterio default of NOT masking and scaling it by default, leaving it as an opt-in for the user, but maybe this can be changed in the xarray's backend too?

Below are examples of multiple ways to open a TIF file with a single band, and the different results:

import xarray as xr
import rioxarray

xr.open_dataset(path).band_data.encoding
{'dtype': 'uint8', 'scale_factor': 1.0, 'add_offset': 0.0, 'grid_mapping': 'spatial_ref', 'preferred_chunks': {'y': 512, 'x': 512, 'band': 1}, 'source': 'myraster.tif', 'rasterio_dtype': 'uint8'}

xr.open_dataset(path, mask_and_scale=False).band_data.encoding
{'grid_mapping': 'spatial_ref', 'preferred_chunks': {'y': 512, 'x': 512, 'band': 1}, 'source': 'myraster.tif', 'rasterio_dtype': 'uint8'}

xr.open_dataset(path, mask_and_scale=True).band_data.encoding
{'dtype': 'uint8', 'scale_factor': 1.0, 'add_offset': 0.0, 'grid_mapping': 'spatial_ref', 'preferred_chunks': {'y': 512, 'x': 512, 'band': 1}, 'source': 'myraster.tif', 'rasterio_dtype': 'uint8'}

rioxarray.open_rasterio(path).encoding
{'grid_mapping': 'spatial_ref', 'preferred_chunks': {'y': 512, 'x': 512, 'band': 1}, 'source': 'myraster.tif', 'rasterio_dtype': 'uint8'}

rioxarray.open_rasterio(path, mask_and_scale=False).encoding
{'grid_mapping': 'spatial_ref', 'preferred_chunks': {'y': 512, 'x': 512, 'band': 1}, 'source': 'myraster.tif', 'rasterio_dtype': 'uint8'}

rioxarray.open_rasterio(path, mask_and_scale=True).encoding
{'dtype': 'uint8', 'scale_factor': 1.0, 'add_offset': 0.0, 'grid_mapping': 'spatial_ref', 'preferred_chunks': {'y': 512, 'x': 512, 'band': 1}, 'source': 'myraster.tif', 'rasterio_dtype': 'uint8'}

Recap:

	xr.open_dataset	rioxarray.open_rasterio
No mask_and_scale provided	'dtype': 'uint8', 'scale_factor': 1.0, 'add_offset': 0.0	🚫
mask_and_scale=False	🚫	🚫
mask_and_scale=True	'dtype': 'uint8', 'scale_factor': 1.0, 'add_offset': 0.0	'dtype': 'uint8', 'scale_factor': 1.0, 'add_offset': 0.0

[snowman2]

It is inconsistent, yet it is consistent.

• xarray.open_dataset defaults to mask_and_scale=True for all engines I am aware of. rioxarray uses the same default to be consistent with the behavior of that interface.
• rioxarray.open_rasterio has always defaulted to false. No change was made for backwards compatibility.