[gdal-dev] Setting NODATA to -nan

[Even Rouault]

> Thanks Even,
> 
> I have submitted <https://trac.osgeo.org/gdal/ticket/6036> with an
> attached TIFF
> infested with `-nan` values.
> 
> While researching this I cam across this question
> <http://stackoverflow.com/q/3772835> which might contain some information of
> interest.

Hum looking more closely at representation of nan values in IEEE754, there is 
at least 12 million different ways of representing nan for a 32bit floating 
point value. And you can indeed have signed NaN... But I'm not sure it is 
really useful to set "-nan" as the declared nodata value (it is stored as a 
string in GeoTIFF) since even 2 NaN values that could be printed in text form 
as "-nan" could have a different binary representation.

The parts of GDAL that test pixel values against nodata have normally a 
special behaviour  to test the pixel values against nodata, when the nodata 
value is one of the NaN values, so I wouldn't expect this apparent 
inconsistency to cause problems

For example when running gdalinfo -stats on your file, I get :

ERROR 1: negative-nan-example.tif, band 1: Failed to compute statistics, no 
valid pixels found in sampling.

Which is expected since all pixels are set to nan (-nan yeah...), so 
ComputeStatistics() actually managed to match the declared nan with the -nan 
as pixel values. (which is quite ironical since a property of nan is that nan 
!= nan, but here nan ~= -nan ;-) )

Did you run into particular problems beyond this apparent mismatch ?

Even

-- 
Spatialys - Geospatial professional services
http://www.spatialys.com