Anna and Yannick opened by stating that the focus of the meeting was to discuss how to partition up the functionality (of sfc obs ops) keeping in mind that we want to minimize the amount of effort to get there. In one extreme we build many operators for every case and on the other extreme we build one operator to handle all cases. Both of these extremes create too much work and result in a difficult to manage system (due to high complexity). What we are after is the sweet spot in between the extremes that reduces effort and complexity.
Anna asked the different centers to present how they handle surface obs in the context of the H(x) calculation.
EMC
Cory presented the following slides:
Cory has started work on this with recently merged pull requests (feature/atmsfcinterp) into UFO and IODA. These handle interpolation of T, Q, U, V from the lowest model level to 10m for winds and 2m for temperature and moisture. Slide 1 has details.
Slide 2 shows future work of which Xiaoyan Zhang is helping. Xiaoyan mentioned that variable transformations will be important, such as those from log or non-linear space to linear space.
UKMO
David Simonin presented the following slides:
We focused primarily on Slide 1. David mentioned that the UKMO handles variable transforms in an obs preprocessor step (orange boxes on Slide 1). Yannick responded that for JEDI we may want to consider using the QC filtering mechanism for this. Currently JEDI does both pre- and post-filtering steps when doing the H(x) calculation, and we could add a means for providing filtering as a stand-alone application to handle cases such as variable transforms in a manner like that of the UKMO.
NCAR
Yali presented the following slide:
This represents the surface pressure assimilation process using the MPAS model. The filtering step is an ad hoc method to reduce errors by removing obs that would have to be interpolated across a large vertical distance. The height correction is being treated as an obs bias. The process is an intermediate step toward the goal of doing multi-variate height correction.
Stan asked how this compares with what GSI is doing. A discussion ensued and it was determined that someone needs to look through the GSI code and make the comparison. The purpose of this would be to figure out what needs to be configurable parameters with the hope of finding some generalized algorithms/approaches.
GSD
Stan presented the following slides:
Stan mentioned that they are looking at how to partition functionality to fit into the JEDI scheme as shown on Slide 2. He noted that GSI is lacking some surface obs functionality and the slides show their thinking on how to fill in what's missing into JEDI. Stan emphasized the importance of getting PBL profiles from the model so quantities such as lapse rate can be distinguishable between the PBL and the free atmosphere. Stan also mentioned that it is important to handle vertically staggered model grid data properly.
NRL
Sarah and Nancy presented the following slides:
NRL has a large variety of surface obs that they assimilate. See slides for details. Two topics came up related to their handling of the obs data which were super obbing and thinning. A discussion on these topics ensued. Yannick noted that we should do both of these functions in UFO whereas IODA's responsibility is to manage and provide obs data and as such should not be altering obs data. We have already implemented a thinning filter in UFO, but since super obbing changes obs values it would be a little more tricky to implement. David added that at the UKMO they create a new obs type that holds the super obbed data which allows the original observations to also be available.
General Discussion
Anna mentioned that horizontal interpolation that spans land-ocean interfaces may require a larger set of points beside the typical four model grid locations.
The question of where height correction belongs came up. Three options were presented
- Filter
- Obs operator
- Bias
Yannick mentioned that he is leaning toward the obs operator approach, but this still needs to be thought through some more. David asked if in the obs operator approach, could different methods of height correction be accommodated. Yannick responded that it might be okay as long as we end up with a small set of operators.
Nancy (NRL) mentioned that it is important to decide what obs to keep around for verification purposes.
Summary
Points for further discussion:
- Where do variable transformations fit in the flow
- It's important to get PBL profiles from the model
- How do we properly handle vertically staggered grid data
- What obs do you keep around for verification
- How should height correction be implemented
- Is it possible to generalize height correction methods
- What needs to be configurable parameters
Action item:
- Compare NCAR surface obs methodology with what is done in GSI
- Yield insight on height correction questions (above)