Hi,

Rasterio's rio-convert program can linearly scale pixels and apply offsets:

$ rio convert --help
Usage: rio convert [OPTIONS] INPUTS... OUTPUT

  Copy and convert raster datasets to other data types and formats.

  Data values may be linearly scaled when copying by using the --scale-ratio
  and --scale-offset options. Destination raster values are calculated as

    dst = scale_ratio * src + scale_offset

  For example, to scale uint16 data with an actual range of 0-4095 to 0-255
  as uint8:

    $ rio convert in16.tif out8.tif --dtype uint8 --scale-ratio 0.0625

  Format specific creation options may also be passed using --co. To tile a
  new GeoTIFF output file, do the following.

    --co tiled=true --co blockxsize=256 --co blockysize=256

  To compress it using the LZW method, add

    --co compress=LZW

Options:
  -o, --output PATH               Path to output file (optional alternative to
                                  a positional arg).
  -f, --format, --driver TEXT     Output format driver
  -t, --dtype [ubyte|uint8|uint16|int16|uint32|int32|float32|float64]
                                  Output data type.
  --scale-ratio FLOAT             Source to destination scaling ratio.
  --scale-offset FLOAT            Source to destination scaling offset.
  --rgb                           Set RGB photometric interpretation.
  --co, --profile NAME=VALUE      Driver specific creation options.See the
                                  documentation for the selected output driver
                                  for more information.
  --help                          Show this message and exit.

Different forms of scaling are possible if you use the Python API, numpy, and scikit-image.

Hope this helps,

Hi,

I apologize for being slow to reply, I've been occupied with fixing bugs in the last couple of releases.

In theory it is possible to add an option to the read method which will cast the data because GDAL already supports this well in versions 2 and up. I haven't pursued this because converting unsigned int to float is trivial, if not cheap, with numpy:

   data = dataset.read().astype('float64')

And if you pass masked=True to read, you'll get a masked array and then can apply a fill_value of numpy.nan under the invalid data mask. Is this useful for you or have you already been down this path?

GDAL's data type casting would not produce an extra copy of the data and so would be optimal for some applications.