s.obsout and moddim about divand.jl

[jmbeckers]

To exclude a possible explanation for the problem can you use alphabc=0 in the parameters and check if the problem remains ?
Thanks

[jmbeckers]

To exclude a possible explanation for the problem can you use alphabc=0 in the parameters and check if the problem remains ?
Thanks

[JianghuiDu]

I wonder if its a problem with sparse_interp

I = localize_separable_grid((eNdint.lon,eNdint.lat,eNdint.w_depth), mask3d, grid)

H, out = sparse_interp(mask3d, I, [true,false,false])

Here I plot out (within the red rectangular, the others are masked because of bathymetry). Even with iscyclic = [true, false, false], the observations are still masked out. I read the code and I didn't find any mention of moddim there.

[jmbeckers]

I tried on a simple case

if the data point is in the domain (not taking into account periodicity) everything is ok

if the data points is out of the domain (say -10 if you define your grid between 0 and 360) then it is not taken into account.

So what is the grid definition ?

[JianghuiDu]

lon = 2.0:4.0:358; lat = -88.0:4.0:88.0
The lon of the masked data are between 358 and 360 OR between 0 and 2. My lon data are all in the range of 0~360. There's no negative numbers.
This plot is better because now I have removed bathymetry mask.

[JianghuiDu]

Maybe its related to localize_separable_grid which appears to have a strange behavior. It returns negative numbers for observations as long as only one dimension is outside the grid range. When calling I in sparse_interp, setting iscyclic = true only modifies that dimension. For example, if an observation is outside lon grid but still within lat and depth grids, it gets a negative I in all 3 dimension, and in sparse_interp setting lon as cyclic only modifies the I in lon dimension, while the other two dimensions are still negative. Shouldn't in such a case all three dimensions be modified to bring the point inside?

[Alexander-Barth]

Thanks a lot for your analysis. This is exactly the issue. I am now testing a fix for this problem.

[Alexander-Barth]

Can you try the new master branch?

[JianghuiDu]

Yes, It's fixed. Thanks!

[JianghuiDu]

A related question. Are the observations outside the grid domain used in the analysis?

[jmbeckers]

Observations outside of the grid are not used for the analysis (neither are observations that fall on land defined by the mask). What can happen is that some of the auxiliary tools doing statistics on the data might not check if the points are on the domain or not but the analysis itself excludes the points falling outside.