We set slope classes so as to have an empty mask, thus 0 valid inlier points for one of the classes :

slopes [%] : [0, 30, 180, 100000, inf]
slopes [deg] : [0.0, 16.69924423399362, 60.94539590092286, 89.94270423958551, inf]

The masks are definied according to those classes :

We check the statistics for each mask with : .get_stats(inlier_mask=mask)
For the last class, no error is raised. We get 0 valid inlier count, 0.0% inlier points and -- value for most of the statistics and nan for 90th percenile and LE90.

@vschaffn @rhugonnet, valentine (i can't @ you) -- means masked, do we work with this type or nan ? :)

-- value for most of the statistics and nan for 90th percenile and LE90.

I think that we need to be consistent. In my opinion, nan could be a good choice for all the statistics.

I agree with nan for 0 valid values in the mask, more widely supported and understood.

@adebardo @quinto-clara @rhugonnet @belletva Thanks for your feedback. I set all the stats to NaN when the raster is fully masked, and added a warning in the logs indicating that the raster is empty.