Dear Joao,

I know that both kind of data can be assimilated in oceanographic analyses. We have the same problem (small domain--> few data). However, also depending I think by your model characteristics, I think that the number of the data is not so crucial for the assimilation system, especially if it is intented for long-term analyses production and not for operational purposes (initialization of fcsts). Maybe you could try at first a different description of your model error (even increasing it artificially just to see if something happens).

I did not try to assimilate the gridded products for several different reasons: gridded data are a ultra-processed product in my view far light-years from the original satellite observations; further they represents a wide temporal window because of the way they are created (1 week or more), so with my 3dvar approach I could have many difficulties to compute exact-time misfits with simulation and also in applying innovations correctly; I do not know papers comparing the assimilation of AT vs Gridded. I think the general choise is assimilate the AT data for the reasons I mentioned here above. Take in mind that on a long run, depending on your system specification characteristics and parameterization, the corrections should be partially accumulate on your final solution as the ocean has a good memory. For sure this is not the case if you have a systematic error or bias in your system that tends to reappear rapidly every time once the corrections was applied.

In any case if you will try assimilation of gridded products please let me know your results, cause this is stringly interesting for me!

Cheers,

Antonio

Using gridded data does not add actual information compared to along-track data. To the contrary: gridded products are actually too smooth, the small scales have been removed by the interpolation method. In addition the error covariances of gridded data are full and unknown, while those of along-track data are almost diagonal and simple. Therefore I advocate strongly to use along-track observations. (Of course this also depends on the scheme -- if you use a simple scheme such as nudging, it might be moe effective to go for gridded data, but an advanced scheme can handle sparse data without trouble.)

Yes, I agree with Pierre about the smoothness of gridded SSH data products and unknown errors. Usually, assimilating gridded data underestimate the fronts. In order for the assimilation scheme to see the along track data, we need to use a time convolution of the error covariances so the memory of along track data is maintained (time correlation models). In variational data assimilation (4D-Var), the along track altimeter data is incorporated once since it takes the satellite few seconds to cover the regional domain (a single ocean model time step).

Dear Joao,

I'd suggesting you have a contact with "Vasilis D. Valavanis" in Hellenic Centre for Marine Research, maybe he can help you more.

Good Luck,

Ali

It is indeed an interesting question. Using 3DVAR, I have carried out a variety of experiments on assimilating along-track data and gridded data from AVISO and from a 2DVAR scheme with various decorrelation length scales to constrain smoothness. Honestly, I cannot come to a convincing conclusion yet. Personally, I think that the assimilation of altimetry data into regional models is pretty much a research topic. In my experiments, one specific thing that I saw is that it seems that the along-track data can enhance the impact of other observations like HF radar velocities, kind of 1+1>2 effect.

First of all, I want to thank you all for your comments.

Just to give some more details, we are running an operational forecast system with a 3 days assimilation window. That means we have roughly 2 tracks (Jason 2) per assimilation window.

I completely agree with the worries about the smoothness and errors of the gridded product, as described by Pierre and Hernan. But the fact is, due to its scarsity, the impact of the along track assimilation in our system is very limited. I really need to check on the comment by Zhijin Li, since we also assimilate HF radars.

Also, I will check on the time convolution of the error covariances Hernan !

Another idea would be to assimilate SLA fields from the NCOM results, that assimilates the AVISO product. So, the field would contain information derived from the model dynamics. It does not solve the problem of the smoothness, given the NCON resolution, but at least it would be in dynamical balance with our boundaries, since we use NCOM as OBC.

A derived question is: Given its high resolution, how long it would take for our model to develop the small scale dynamics "dumped" by the NCOM or AVISO assimilation? Is it "applicable" for an operational forecast system?

I will start doing some tests on this direction... and still appreciate any thoughts you would like to share!

In order to get more tracks in any one inversion, you could get longer time windows (which would come back to overlapping windows if you keep the same assimilation frequency). This usually works nicely.

Also make sure that you use all available satellites, not just Jason! Actually this might be the key to improving your solution.

If you assimilate gridded products (AVISO or NCOM), make sure you do not change the small scales not contained in those products. This comes back to forming the innovation from the spatially lowpassed model fields. I think this would work with your 4Dvar, however you might check the spectral contents of your adjoint variables fields at initial time, to make sure that your particular adjoint dynamics did not introduce small scales (otherwise you might try to increase adjoint diffusion).

I see a few familiar names here, hello!! The answer depends on the model and its resolution. For higher resolution models, I would assimilate along-track and up to 3 days worth of tracks at a time for ocean simulations and probably only one day for coastal simulations. For eddy permitting models, I would assimilate an interpolated product. Also make sure that your assimilation scheme is giving enough weight to the observations so that they do make a difference. The following is just a guess: try assimilating daily tracks all day long in the model, not just at one time, and interpolate in time between the daily tracks. cheers, arthur

Hello to everybody!

I'm Andrea Bergamasco and some time ago I worked with Paola Malanotte Rizzoli on the data assimilation using Adjoint Model approach.

I agree with last comment of Arthur Mariano, but also with Pierre De Mey and Hernan Arango…

the problem is not easy, because you need both horizontal scale signals, not only the along track one. In the past , when I was involved in this kind of business, I used a decorrelation time in order to try an "smooth" interpolation between tracks.

Cheers

Hi Andrea;

Thank you for your comment!

But, if you use a decorrelation time wouldn't you be incurring in the same problem as using the AVISO gridded product? That is, to introduce an "artificial" information into the model...

By the way, we are now running some assimilation tests using both along track and HYCOM derived SLAs. We basically have only Jason-2 in the model domain (looked for Envisat and Jason-1, and could not get access to HY-2), every 2-3 days. We "repeat" each satellite track during a whole day.

Am I forgetting some other altimeter that is still active?

Thank you for all the constructive comments;

To give you some feedback on my progress on this issue, I attached a figure showing the results from 2 experiments, one with assimilation of alontrack SSH from Jason 1 and 2, and other with the assimilation of this same data plus the HYCOM SSH outputs for the same period. This HYCOM simulation was conducted with assimilation of the AVISO gridded product, and is used to give the boundary conditions to the ROMS simulations. The figures show the alongtrack observations minus the ROMS results. As you can see, the inclusion of HYCOM leaded to better comparisons with the alongtrack SSH.

Thank you for the update Joao. If I understood well in the experiment #2 you assimilates with ROMS the output of the coarse model that also provides boundaries for ROMS and that in turn assimilates gridded product. 1 question and 1 observation:

-Do you think that the improvement would be due to the larger amount (indipendently on the tipology of the source) of data?

-It could be that validating with AT the ROMS output that assimilates HYCOM (in turn assimilating gridded data ....produced from AT processing!) could implie that the ROMS output and the AT data used for validation are no longer indipendent and that some autocorrelation is introduced....but I'm not sure of course. This could cause a better result in validation.

Hi Antonio;

Yes, definitly is the quantity of data. While the along-track data provide roughly one pass every 2 days, the HYCOM outputs cover the whole domain every day. It is important to emphasize that we defined higher errors for the HYCOM SSH assimilation (~2 times the along track errors).

In relation to your comment: It is really interesting that the RMSD of the ROMS results in relation to the along-track data gets smaller when we add the HYCOM results assimilation. And the RMSD of the HYCOM results in relation to the along track data is 2 times the magnitude of the RMSD of the ROMS results (without HYCOM assim.) in relation to the same data.

I still have to "think through" these results...