[gdal-dev] Prevent extrapolating with ReadAsArray when resampling

[Even Rouault]

Rutger,

ReadAsArray() / IRasterIO() with non-nearest resampling use the interpolation code of 
overview building (except for Lanczos where it uses the warping code), which was of course 
targetted at downsampling rather than upsampling. Nodata is taken into account when 
mixing together source pixels for a target pixel (those pixels are assigned a zero weight in 
convolution computations), but it seems that we should likely skip entirely a target pixel if 
the center source pixel corresponding to it is nodata, so that nodata areas do not fade 
away (which is done in the warping code). What surprises me is that you get different 
results depending on the nodata value (I could also see that)

Regarding Lanczos, might be a bug specific to it indeed. Didn't look.

Perhaps file this as a ticket about this.

Even

> Dear list,
> 
> I regularly use (and enjoy!) the resampling option when using the
> ReadAsArray method of a Dataset. However, i noticed that in the case of
> nodata values in the raster, the method extrapolates "into" those nodata
> values. This is not always desirable in my case, especially when using any
> resampling algorithms other than Bilinear or Nearest, since these can
> overshoot significantly.
> 
> This made me wondering if there is a way to prevent this extrapolation? One
> thing i noticed it that when the nodata value is non-numerical, like NaN,
> there is a lot less extrapolation (eg more nodata values returned). That
> could be used a workaround, but it not convenient to map numerical nodata to
> NaN values before reading the input.
> 
> In the process of exploring the behavior, i also encountered some issues
> (perhaps a bug?) when using Lanczos in combination with certain datasets and
> certain output grids. This happened especially when the output size is an
> uneven factor of the input size, like (buf_xsize=xsize*3,
> buf_ysize=ysize*3) or (buf_xsize=xs*5, buf_ysize=ys*5) etc.
> 
> Here is a notebook showing what i did with a linearly increasing grid:
> https://gist.github.com/RutgerK/30fbd03d9b2ab79861eedfbb95064732
> 
> And a random grid:
> https://gist.github.com/RutgerK/bdbd6c124224cf49961082e1620ce604
> 
> Regards,
> Rutger
> 
> 
> 
> 
> 
> --
> Sent from: http://osgeo-org.1560.x6.nabble.com/GDAL-Dev-f3742093.html
> _______________________________________________
> gdal-dev mailing list
> gdal-dev at lists.osgeo.org
> https://lists.osgeo.org/mailman/listinfo/gdal-dev


-- 
Spatialys - Geospatial professional services
http://www.spatialys.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/gdal-dev/attachments/20180105/d2886b36/attachment.html>