Improvements for the Objective Analysis codes

current suggestions

1. BUG: If the parameter coord is reset from the GRIDS netCDF file, the method of calculating the correlation function (cartesian/spherical) is not reset.
   • If not the method, then something isn't properly reset.

2. Grab the code/MATLAB routine someone has (CJL) to estimate correlation parameters from data.

3. Keep current with any changes to ASCII data format definition
   • Select variables by name.
   • Pay attention to fill values.
   • Allowing velocities will require V-grid/T-grid distinctions.

4. DOCUMENTATION
   • Need references like these.
   • need description of strategy for selecting vertical levels. A simple statement about resolving the vertical structure of the data should suffice. Or maybe a short paragraph.

5. Need some method of handling instrument expected error. CTDs generally more accurate than XCTDs or XBTs and climatology generally less accurate than any other. This should somehow be used.

6. Stability.
   • From Reiner:

     thank you for mailing me the AFMISOPS packages. It works well - all
     profiles are stabilized after using it, 
     
     but
     ---
     
     the oag output is still gravitationally unstable, and I suppose I have
     found out, why: The instabilities are produced by the OA acting on
     climatological T and S profiles of different vertical extent. 
     
     In order to understand what I mean, please run the attached short MATLAB
     script STAB.M. It illustrates what the OA might produce from input
     profiles of different vertical extent.
     - green: long (> 2000 m) profile , stable
     - blue: neighbored short (< 2000 m) profile, stable
     - red: OAed profile (in this case only the mean profile) located in the 
       neighborhood of the green and blue profile ---> unstable below
       2000 m! This is because below 2000 m the OA "sees" only the green
       profile.
     
     I suppose that there are only two methods to overcome this problem:
     1. Stabilizing the OAG output
     or
     2. performing the OA on potential density surfaces instead of
        horizontal. But before doing that, it has to be checked, to 
        what degree cabbeling effects might create instabilities as well.
     

7. OAG: Get the conjugate gradient on-line. Don't write it assuming the full matrix exists.

8. Perhaps want to rework OA(G) so all data isn't stored at the outset. Either grab only needed data or only read in one field at a time.

9. Look into using SVD in local OA. Possibly could use this to replace the arbitrary RAVG & TAVG parameters used to ensure linear independence.

10. Compare OAG's method of melding the mean & synoptic fields with its method for melding the climo & data mean fields. If different, consider making uniform. Consider translating that algorithm to the local OA.

11. Allow the OA(G) to read the mean field. Either in an existing OA(G) output file (like GRIDS) or some other means.

12. Enable the masked correlation.
    • Remove "masked" observations.
      • Or:
        1. find straight distance from "masked" observation to coastline.
        2. start from that coastline point and find distance to mapping point.
        3. add the above.
    • Decide on logic for profiles external to grid.
    • Modify PE_mask to not remove last exterior coastline point.
      • Perhaps a separate set of OA(G) coastlines which bound all land like islands.
    • Look in the directory /home/alboran2/work/haley/OA_Dist for the last attempt.

Blasts from the past

1. "The diagnostics of the observed field with flagged (-9999) values is not handled the same in OA and OAG." Check this out.

2. (11/94, PFJL) getobs.F:
   • "I think there is an error in the time conversion from (jday,ihour) to dfrac"

3. (11/95, PFJL) In OA code, using idhopt=4 will yield different results if done with one level or many.

4. (11/95, PFJL) Also concerned that OA(G) codes extend horizontally. Might give unreasonable results for points beneath topography. This could be a factor when PE_initial has to interpolate.

Really "off the cuff" stuff

1. Instead of interpolating data in the vertical, consider optionally averaging data in boxes.
   • Don't use PE type boxes. Probably better to average from halfway-to-the-level-above down to halfway-to-the-level-below.
   • Leave as an option, it assumes that enough levels are being analyzed to represent the water column. It wouldn't be a good idea for just checking a couple of depths (e.g. surface and main thermocline).