Erroneous values in destination image with gdalwarp's interpolations

[andrieu]

Hello,

I have been trying to use gdalwarp to make some image merging and reprojection.

Unfortunately, with gdalwarp, the use of interpolations others than bilinear will produce images with wrong values.

My source pictures are Byte, 1 band, nodata = 0.

I have trying to use various destination datatype but they all lead to the same results : values in the destination image are below 0 (I even found a value around -80...) and so are considered as nodata.

My conversion script is :

gdal_translate -projwin 713676 4783608 713692 4783597 PIC_SS_P_821_8bit.tif little.tif

gdalwarp -t_srs EPSG:4326 -r bilinear -ot Int32 -dstnodata -100 little.tif little_bilinear_1.tif
gdalwarp -t_srs EPSG:4326 -r cubic -ot Int32 -dstnodata -100 little.tif little_cubic_1.tif
gdalwarp -t_srs EPSG:4326 -r cubicspline -ot Int32 -dstnodata -100 little.tif little_cubicspline_1.tif
gdalwarp -t_srs EPSG:4326 -r lanczos -ot Int32 -dstnodata -100 little.tif little_lanczos_1.tif

I then use a gdal_translate to add compression, tiling and correct nodata :

gdal_translate -ot Byte -a_nodata 0 -scale 0 255 1 255 -co TILED=YES -co COMPRESS=DEFLATE -co PREDICTOR=1 source dest

And finally a gdaladdo pass.

The reason I make a gdalwarp in a signed datatype is that the interpolation does not take the nodata value in consideration so my destination image will be filled with wrong nodata. So to avoid that, use of gdalwarp in signed datatype followed by a gdal_translate in unsigned datatype with scaling will produce correct nodata.

By the way, I am using FWTools 2.2.9 and 2.4.3 (same results).

This is very sad because image quality is better with lanczos than with bilinear...

Regards,

Benoit Andrieu

[andrieu]

Source image

[andrieu]

Bilinear result with Int32

[andrieu]

Cubic result with Int32

[andrieu]

Cubicspline result with Int32

[andrieu]

Lanczos result with Int32

[Even Rouault]

Several things :

• if you use a negative nodata value, yes, you must use a signed datatype. I've added in r17658 a check to clamp nodata values if they go outside the valid range of the datatype.
• the negative values you get for cubic, cubicspline and lanczos are not surprising, considering the non-linear underlying maths. For example, on your example with cubic, at a moment we compute CubicConvolution(d, d*d, d*d*d, f0, f1, f2, f3) where d = 0.136, f0 = 37, f1 = f2 = f3 = 1. The result is -2.66. bilinear interpolation of course doesn't allow for those undershoot or overshoot effects.
• in those cases, using an unsigned datatype leads to truncation to positive range, hence those negative interpolated values are clamped to 0. If 0 is the chosen nodata value, then it might be a problem. So in r17659, I've added a check so that if a computed pixel value is equal to a destination nodata value, we use the integer above or under instead.

[andrieu]

Thanks for the changes ! That should do the trick !!! I'll test those as soon as possible (I'm on holidays...)

[andrieu]

Ok, as you said, computed values equal to nodata are set to the above value.

Thanks !