[gdal-dev] Bug? Gdalwarp produces images with dark blocks at certain scale

[Michal Migurski]

I think I'm experiencing a gdalwarp bug. Maybe two gdal bugs? Dark-colored blocks and misregistered output from a VRT with about a dozen or so ground control points. The problem looks like this:
	http://mike.teczno.com/img/gdalwarp-darkness/landcover-1km-merc-z7.jpg
	http://mike.teczno.com/img/gdalwarp-darkness/landcover-tiles.jpg

Sample files linked below.

Possible Bug #1: Dark blocks in gdalwarp output.

I have a TIFF of landcover for North America, downloaded from nationalatlas.gov. The original is in a Lambert projection that's slightly off by a few km and I need it in spherical mercator for tiling, so I've created a VRT file with ground control points expressed in the projection of the original.

Original data from here:
	http://nationalatlas.gov/mld/landcvi.html

Both of my files linked here:
	http://mike.teczno.com/img/gdalwarp-darkness/landcover-1km-rgb.tif (8MB TIFF)
	http://mike.teczno.com/img/gdalwarp-darkness/landcover-1km-to-merc.vrt

When I convert the VRT to tiles using GDAL's python bindings, I get darkened tiles but only at one specific scale that's close to the original scale of the image. The output looks like this at zoom=7:
	http://mike.teczno.com/img/gdalwarp-darkness/landcover-tiles.jpg

The color in the top-left corner of the image is correct, while the color in the bottom right corner of the image is darkened but fades out to normal as you pan south and east. If you zoom out to z6 or in to z8, the problem disappears - it exists *only* at z7, a scale very close to the native scale of the source image.

I thought that I could get around all this by pre-projecting the image in one shot to a spherical mercator TIFF and then having the tiles cut from that image. When I warp it to an image at zoom=7, the dark, block-shaped areas appear again:
	http://mike.teczno.com/img/gdalwarp-darkness/landcover-1km-merc-z7.jpg

Warping to z8 or z6 does not show this problem.

I used these commands to do the warp:
	gdalwarp -t_srs EPSG:900913 -te -18701674 2264982 -5788613 10750786 -tr 1222.992 1222.992 -r cubicspline -tps   landcover-1km-to-merc.vrt   z7.tif
	gdalwarp -t_srs EPSG:900913 -te -18701674 2264982 -5788613 10750786 -tr 611.496 611.496 -r cubicspline -tps   landcover-1km-to-merc.vrt   z8.tif

Possible Bug #2: Misregistered gdalwarp output from VRT.

I figured that I could cut tiles from a non-blocky, zoom=8 spherical mercator image. Thanks to the ground control points in the VRT and visual evidence from the tiles I've been working with, I assumed that simply running gdalwarp would result in a correctly-registered output file. However, when I took the zoom=8 TIFF generated above and tiled it, I was seeing the area around San Francisco projected too far north by about 20km, and the area around New York too far south by about 10-20km. The whole map appeared to have been rotated or distorted. Why would gdalwarp be unable to create a correctly-registered image that the python bindings seemingly had no trouble with?

I realize this is a big question, and I hope that the various provided files help.

-mike.

----------------------------------------------------------------
michal migurski- mike at stamen.com
                 415.558.1610