Fail on 5 and 10 km grid

[mjaehn]

@jonasjucker I tried out both R2B8 and R2B9 grids but they fail. I thought at least the 10 km would go through. Or is there another reason?

[mjaehn]

DOM_1_extpar.log
DOM_1_extpar_ndvi_to_buffer.log

[drieg]

@mjaehn @jonasjucker these failures seem to be a problem in the conservative remapping algorithm of the latest CDO releases. CDO-2.0.5 still works and I got feedback from the CDO developers that it is also fixed in the development version. So the next release might resolve this.

It seems to happen when single grid points of the source and target grids are collocated. The chances of this to happen in combination with a regular lon/lat grid are higher for R2Bk grids than for R3Bk grids since the points are located on circles of latitude.

Currently possible workarounds

• Rotate the grid slightly to avoid the collocation of grid points. E.g. Global R2B8, rotated by 36deg #62 rotates by 36 degrees which is also done operationally at DWD to avoid a pentagon point over Europe which is half the distance between two pentagon points (72 degrees).
• Use cdo-2.0.5 until a new release fixes the issue
• Use a different interpolation algorithm. Is bilinear an alternative?

[drieg]

Use cdo-2.0.5 until a new release fixes the issue

I received some additional information: The old versions simply do not perform a check which can lead to erroneous results. In the development version (i.e. next release), the fix is to increase the tolerance by a factor of 10. While the R2B8 grid does not run into this error anymore, there might still be other configurations that will encounter this. So this is no perfect solution

[mjaehn]

Thanks @drieg for pointing out the source of this problem!

So next cdo release is 2.4.5? Would it make sense to rotate the grid by default?

@jonasjucker Can we try out bilinear remapping?

[drieg]

Would it make sense to rotate the grid by default?

You can still run into the same issue, so I think the answer is no.

More generally, anyone creating a grid should think about rotating the grid to avoid pentagon points in the target region. But this is really up to the user and their specified target region. I don't think that an algorithm can cover this in a satisfying way since it requires to know the preferences of the user

[jonasjucker]

Honestly I want to limit the amount of work put into fixing Extpar by bugs discoverd by extpar-request to a minimum.
So my fix will be to lock cdo to a version that does not show the crash.
To adress this issue in Exptar at some point it makes sense to open an issue there.