Analysis Increment of JEDI LGETKF

Hi all.

I am testing LGETKF of public JEDI repo and have questions of analysis increment of LGETKF. I employed ‘tutorial_case data’ (c12 global grid, geos, 2018-04-15_00z) with single scatwind (surface) obs with two different ‘fraction of retained variance’ parameters (0 and 0.9).

I checked vertical profiles of the nearest grid-point where obs exists. When I checked difference of ‘ua’ in ensemble mean between analysis and background, increments seems to be acceptable in that they are mainly located in the bottom levels and the one (frac0.9) which I put fraction as 0.9 seems to have more restricted increments around level 25 (frac0 shows increments up to around level 30). I think this seem to be related implicit model-space vertical localization using modulated ensembles.

However, when I checked the difference of ‘ua’ in first ensemble member (001), there are things which I can’t understand. You can see that both have analysis increment at all vertical levels and there is a peak at higher levels. Since I employ surface-level observation, I expected that there will be more changes at lower levels. In addition, you can see that the magnitudes are different between ensemble mean and member 001. Since there is little change in ensemble mean at upper levels but first ensemble member has significant difference. It seems to me that the analysis perturbations at upper levels are reorganized to have the same (or similar) ensemble mean but have different perturbations from background one.

Judging from the code, it’s expected to see the changes in upper levels in case of LGETKF when compared to LETKF. However, I think this is too much of a change for ensemble member when model-space vertical localization is considered. Do you think this is a expected result?


Hi Jun, You are correct at evaluating the first plot. You can clearly see the impact of localization. As for the second plot, what you see is the ensemble members having large perturbations aloft (which makes sense since the magnitude of the winds are large aloft). Part of the work done by the ETKF is rotating the ensemble members to preserve orthogonality of the ensemble members (unlike breading methods, which collapse all members on the fastest growing mode). So the difference between the red and blue lines is just the rotation of members with no impact on the mean or ensemble std. Hope this helps.

Hi, Sergey.

Thanks for reply. Your explanation helped me understand the result better but I still want to know why this result was obtained (Actually, my interest is not the difference between FRAC0 and FRAC0.9 but the analysis increment of LGETKF on member 001) . I will investigate this further by comparing the result of LETKF and checking inflation options. I will update this post after I get the results.