Ensemble strangely narrow despite perturbations

[cbachand]

Hi, I am working on generating an ensemble using Noah-MP forced by MERRA2. How exactly do additive and multiplicative perturbations work on LIS? I've specified very high multiplicative perturbations for rainfall rate and snow depth (up to 4), but my ensemble is still surprisingly narrow. Thank you!

[karsenau]

Hi @cbachand:

Would you be able to share with us your two met forcing attribute tables, including the perturbation attribute table?

Thanks!
Kristi

[cbachand]

Hi Kristi, here are those files. Thank you!
forcing_atts.txt
prog_perts.txt
prog_atts.txt
forcing_perts.txt

[karsenau]

Hi @cbachand:

Thank you for sharing. I noticed that your "tcorr" values have a value of every 3-days for your forcing perturbations file. Have you tried shorter time-correlation values, e.g., daily = 86400?

Also, in your forcing_atts.txt file, you may want to increase your "varmax" values by a little bit, e.g., shortwave radiation, you could try 1400. and rainfall rate, you could bump that up to 0.01 or 0.1 even.

For your snow state attributes files, I might recommend making the "varmax" values as reals. As for the perturbation file, I would recommend decreasing your "tcorr" values here (e.g., try daily or subdaily values) and, for now, turning off the "xcorr" and "ycorr" values (set to 0 for testing). Also, be careful of setting your "std" values too high for certain state variables, like snow.

Hope this helps in seeing a little more spread in your forcing and state values. If you are learning to "tune" these parameters, it would be helpful to review some key papers out there on this exercise.

[cbachand]

Thank you for the quick response. What do you mean by setting the "varmax" values as reals?

[karsenau]

Your varmax values in the "prog_atts.txt" file are set as integer entries (e.g., 100, 3000), but we typically enter those as 100.0 or 3000.0.

[cbachand]

I made the changes you recommended (see my files below), but my ensemble is still very narrow. For reference, here are plots of ensemble mean snow depth and the ensemble standard deviation, both averaged over time. I am using a multiplicative perturbations for snow depth of 1.5, so I would expect the ensemble spread to be similar to the total snow depth, if not greater. Why is this not the case? Why is my ensemble standard deviation still so low?

forcing_atts.txt
forcing_perts.txt
prog_atts.txt
prog_perts.txt

[dmocko]

Hi @cbachand

I am no expert on setting these values or how much ensemble spread we might expect.

But if you haven't already looked, you can also see some forcing and model prognostic attribute files in this directory:

https://github.com/NASA-LIS/LISF/tree/master/lis/configs/557WW-7.4-FOC/NRT_GLOBAL/attribs

See the "forcing_" and the "noahmp_snow_" attribs and pertattribs files.

David

[karsenau]

@cbachand:

You can take a look at the attribute files that David (@dmocko) pointed you to on our LISF GitHub master branch. You can see some of the typical values that we work with and options selected for forcing fields and NoahMP snow states.

One note on the temporal averages that you are showing in your two shared plots: The range of std dev values for snow depth that you are showing actually look reasonable for the amount of snow depth for your area of interest. You want to make sure that your ensemble spread isn't too great, like not being at or greater than your total snow depth. Large std dev and ensemble spread values would indicate more uncertainty for the model states. Depending on what you estimate your DA obs std dev to be (also dependent on the observation type), you want to give representative std error values that don't overestimate the "weight" of the observations in your analysis updates.

[cbachand]

Hi, it looks like my perturbations to prognostic variables actually aren't working at all. I ran LIS with two different prognostic perturbation stds, and the two model runs yielded identical results for snow depth and SWE. Is there a mistake in my input files (see below)?
prog_atts.txt
lis.config_ens.txt
model_output.txt
prog_perts.txt
large_prog_perts.txt

[karsenau]

Hi @cbachand:

See if you can increase your frequency of which the perturbations are applied:

State perturbation frequency: "12hr" ==> e.g., "3hr" or "6hr"

And for "tcorr" in your prog_perts.txt file, see about reducing the time for the correlation, e.g., from 43200 to 10800. Does these changes have any effect on your pertubations?

In future runs, we also recommend applying the following option to be turned on for perturbation bias correction:
Apply perturbation bias correction: 1

to reduce having the ensemble be impacted with lognormal or multiplicative perturbations.

[karsenau]

Also, I noticed in your lis.config file that you are running with a 100 members per tile. Have your tried various ensemble sizes (e.g., start with 20, then 50, etc.)? Not that this should matter a lot but was curious how the distribution and spread change at all with size applied.

[cbachand]

Does the model time step affect the perturbations?

[karsenau]

It might -- if going from a 15-minute vs. 1-hour model timestep, the snow-soil moisture profile matrix updates may have slightly different timings with different intervals. I haven't tested different configurations of those, so you might have to test and see. In the past, I have typically run Noah LSMs with a 15-minute timestep, but other studies, e.g., using LIS/WRF-Hydro, have used as coarse as 1-hour. You could look to see if any of those studies have looked at that and any impacts on the perturbations.