Non-standard affine transforms handled incorrectly

[perrygeo]

If you have a dataset with inverted y axis (i.e. the affine.e parameter is positive) a lot of assumptions are broken since North is no longer "up".

The bounds_window must return row_start < row_stop due to rasterio design. So why not internally convert to north-up on read? The window_bounds functions always return w,s,e,n and we'd just need a few conditionals to ensure new_array was a north-up array.

Work started in the southup branch

[kmoses45]

Is this what could be causing 'ValueError: negative dimensions are not allowed' in zonal_stats(vector, raster, categorical=True)?

[perrygeo]

Possibly. Try rio info on the raster and check the transform.

[kmoses45]

"transform":[1.0, 0.0, 0.0, 0.0, 1.0, 0.0]
Both vector and raster are in EPSG:4326

[perrygeo]

With that transform, it's likely that your data lacks georeferencing according to rasterio/GDAL. That's the identity transform and is unlikely the true transform for your data unless you have exactly 1 degree cells with the upper left corner at lat 0, lon 0.

EDIT: To your original question re: ValueError: negative dimensions are not allowed - yes that is currently the behavior when the y axis is inverted (i.e. the 5th element of the transform is positive)

[kmoses45]

Ah, that's it! I was using an .ovr file that was georeferenced in Arc. Converted to .tif, georeferenced, and now it works. Thanks for the help. Rasterstats is an awesome module!

[pierstitus]

Good to see you know about this problem, it took me a while to find this issue.

It would be nice to do a little check on affine and give a warning when e is positive (or a is negative, even though that's way less common)

For those working with ndarrays, here a little function to flip the raster and affine matrix (assuming coordinates are on center of grid cells):

def flipud(raster, affine):
    raster = np.flipud(raster)
    affine = Affine(affine.a, affine.b, affine.c,
                    affine.d, -1 * affine.e, affine.f + (affine.e * (data.shape[0]-1))
    return raster, affine

Thanks for the good work!

Would you like me to work on this issue?

[raybellwaves]

FYI the above function has a couple typos (missing a bracket and data not defined). here's the fixed version:

def flipud(raster, affine):
    raster = np.flipud(raster)
    affine = Affine(
        affine.a,
        affine.b,
        affine.c,
        affine.d,
        -1 * affine.e,
        affine.f + (affine.e * (raster.shape[0] - 1)),
    )
    return raster, affine

[AJSchiller]

I have an idea for this. Can anyone share vectors and raster data that do work so that I can make sure I don't break anything?