Review of “On Process-Oriented Conditional Targeted Covariance Inflation (TCI) for 3D-Volume Radar Data Assimilation” by Vobig, Potthast, and Stephan

By Altug Aksoy (University of Miami/CIMAS and NOAA/AOML/HRD)

24 October 2024

Recommendation: Major revision

Synopsis: This manuscript investigates the impact of a new targeted covariance inflation technique that attempts to adaptively inflate the ensemble background covariances to capture convection in its early stages when the ensemble systematically lacks convection where reflectivity observations suggest that convection is being realized. Overall, I find this approach very promising and its specific application to target early-stage convection makes great sense. Other than some “systematic” grammar issues, the manuscript is well written and easy to follow. I have some major comments that the authors should address carefully and satisfactorily, but otherwise I expect that the manuscript will ultimately be a useful contribution to the field.

Major comments:

(1) On the general application of the targeted covariance inflation (TCI) method, as explained in section 3.2: (a) Can the authors provide more information about whether they have performed any sensitivity experiments when determining the individual conditional check thresholds explained through equations 13-17? It would be helpful to understand how sensitive this method is to the parameter values chosen, and thus, how sensitive the DA performance is to these choices. Please also see my detailed comment below regarding L334-336. (b) Further question/clarification is needed on the inflation of the ensemble members as explained in equation (18). I assume that there is some kind of random perturbation generated for each ensemble member as a result of the TCI algorithm. I see the mentioning of a mean and spread on L134, but I can’t find these repeated anywhere else. Can you please explain this better? (c) Finally, the authors mention on L250-255 that while the default reflectivity observation error is 10 dBZ, when the TCI algorithm is applied, it is reduced to 2 dBZ so that those observations are given more weight compared to the background. I think again, it would be useful to provide more information on whether there were any sensitivity tests to come up with these values and how sensitive the DA performance was on these choices. (My detailed comment regarding L334-336 is also valid here.)

(2) As I was reading the description of the individual cases in section 4.2, it became clear to me on L336-339 that the cycled nature of this experiment actually obscures the direct impact of the TCI method on convective initiation. It would be very helpful to generate a new experiment where the TCI method is applied in a cold-start scenario. In other words, rather than investigating the impact of TCI on a case where the “background” reflectivity is already influenced by the cycling of TCI in previous cycles, it would be more illuminating to choose a particular initial time and simply apply the TCI on the Control experiment at that particular analysis time. This simple (and cheap) analysis/forecast experiment would provide a more direct and useful picture of how the TCI method impact the generation of reflectivity in forecasts where the original ensemble had very little or no reflectivity to work with.

(3) On the discussion of observation error statistics (section 4.3.1): It is important to caution that not all radiosonde locations should be expected to be directly influenced by the application of the TCI. After all, the TCI is applied only in a very small number of locations in each cycle, and thus it would be statistically rare for the location of radiosondes and TCI application to match. However, since TCI introduces artificially amplified humidity near the height of 4-5 km, it is plausible to expect that such effect to accumulate over time during cycling and spread out to the computational domain. But there are quantitative means to inspect this rather than speculate. For example, if this is the case, the “bias” should grow in time as more humidity bias is accumulated in the cycles. Furthermore, if indeed TCI is introducing these biases, those locations should also be associated with deficient ensemble spread in reflectivity. It would be therefore a more complete investigation if (a) time tendencies are investigated, and (b) more complete observation-space diagnostics are employed to take a closer look at bias, bias-corrected rmse (see Aksoy et al. 2009, MWR) and spread consistency ratio (see Dowell et al. 2004, MWR).

(4) While this manuscript focuses on the situation where the reflectivity observation suggests presence of precipitation, but the underlying model background has none, I couldn’t help but wonder whether the authors are also addressing the opposite situation, i.e. spurious convection in the model background where the observations have none. Aksoy et al. (MWR 2009) and later studies have addressed this issue in various ways and I’m wondering whether the authors’ DA system also addresses this “opposite” problem. From the discussion on L417-423, this doesn’t appear to be the case, but a clearer explanation early on, perhaps even in the introduction, could be a useful background information for the reader to follow these results.

(5) The references are heavily tilted toward older publications and many books. While some of this is justifiable to introduce the basic concepts mentioned in the manuscript, the authors are strongly recommended to include relevant citations that are more recent. The issue of and research on covariance inflation, even in the context of convective forecasting and radar data assimilation, now spans nearly two decades and there are many recent publications in the literature that are very relevant to be cited. I strongly recommend that the authors go through their literature and reduce the number of citations to older manuscripts and books and instead provide a list of newer references that are relevant to the subject matter.

(6) This is not “major” science-wise, but I found the common use of the present continuous tense (rather than the present simple tense) rather unnecessary and even grammatically wrong. To give an example, on L42-47, the uses “is effectively rejecting them” and “is failing to synchronize” are incorrect and should be modified as “effectively rejects them” and “fails to synchronize”. This is because these are general statements and not actions occurring directly in the present time. The authors should revise their manuscript for this type of very frequent misuse.

Minor Comments:

L52-55: Is this issue actually documented in the references mentioned in this paragraph, or are the authors inferring this on their own? If this is documented, please provide references.

L118-120: Can you refer to specific equation numbers so that the reader can follow this conjecture easier?

L120: observations -> reflectivity observations

L131: What does “COSMO” stand for?

L142: variables -> prognostic variables?

Figure 2: Can you please translate the right panel to English? Also, the image quality of the right panel should be improved.

L150: Is the 1-km radial resolution here the actual along-beam measurement resolution or some kind of average?

L151-152: It would be useful here to provide at least the basic forward operator equation that converts model variables to reflectivity. Is it the standard conversion that most studies utilize? (Any other factors mentioned at L155 and later obviously would not be included in this equation.)

L160: Can you please be more specific about whether LHN here is something that is done outside of reflectivity assimilation?

L174: Eliminate the word “precisely” since you already use the word “specifically” in the same sentence.

L176: The model itself is -> This model is

L181 and later: artificial -> artificially

L183: potential negative effects? Can you provide examples here as to what those potential negative effects could be that you’re trying to minimize? Are there examples of this in the literature?

L211-212: This disclaimer is actually more confusing than explanatory to me. This manuscript specifically focuses on the TCI model, so it’s not clear which details of it are not relevant or beyond the scope of the present work. I suggest being either clearer about this or removing this disclaimer if it’s not critical to the description of the TCI model.

L223: for specifying -> to specify

Equations 14-16: Since you seem to have enough space in the left column, can you please spell out these condition checks. Specifically, it’s not clear what “Det.” Here refers to.

L236: What do you mean by “deterministic member”?

L286: Please indicate what “RealPEP” stands for.

Figure 4: Since the authors appear to prioritize the 15-dBZ observed reflectivity threshold here, I realize that it may be more helpful to change the order of the equations 13-17 so that observation check and height check (equations 16-17) are listed before the model-related checks (equations 13-15).

L289 and related: cycles -> experiments? (To me at least, this would be more informative.)

L319: of only a few

L322: It’s not clear here what you mean by “production of reflectivity”. Do you mean “convection initiation”?

L325: Here, “like” and “e.g.” are both meant to provide a list of examples. Therefore, please remove one of the two.

L327: For observing -> To observe

L331-332: The use of the commas in the sentence makes it difficult to follow it. Please revise to make this sentence clearer.

L334-336: Isn’t this somewhat contrary to the way TCI is set up to capture convective initiation itself? When I read the description of TCI, I expected to see better examples of mature convection where otherwise the model did not produce any convection, specifically because TCI was applied during the initiation stage. The figure here shows the state of convection after one hour, which in convective time scales is a long duration to generate mature convection especially considering that typical convective overturning time scale is O(10 min). This takes me back to my major comment (1a and 1c): Could smaller reflectivity threshold values perhaps result in more robust convective initiation that is “caught” earlier in the convective lifecycle? Have you performed any sensitivity experiments along these lines? 

L345: spatial -> horizontal

L346: for referring to -> to refer to

L386 and elsewhere: Please avoid subjective adjectives such as “striking”. What is the reason this result is striking, without any quantitative assessment? Furthermore, if this result is “unsurprising”, how is it “striking” at the same time?

Figure 9: Only looking at the row labels, it wasn’t clear what was meant by “init=summed” and “init=12”. If I understand correctly, these can be replaced by “init = all” and “init = 12Z”, respectively, to make it much easier to follow the panels without having to refer to the figure caption.

L412-413: Please make it clear that “afternoon” here is with respect to the local time around Germany readers from other time zones might be accustomed to associating these Zulu times with their own local times.

L433: is very good in initializing -> successfully initializes

L434: larger -> stronger

L437: Double use of “further”