Please see my comments in the attached pdf.

Review of the manuscript "Improving estimation of a record breaking East Asian dust storm emission with lagged aerosol Ångström Exponent observations'', by Yueming Cheng, Tie Dai, Junji Cao, Daisuke Goto, Jianbing Jin, Teruyuki Nakajima and Guangyu Shi.

This paper assimilates AOT and AE from AERONET measurements for an East Asian dust storm and computes dust emissions analyses. The work is technically well done and the manuscript is relatively well structured. In general, I recommend to address the aspects below before publication.

FR simulation has a large bias. It is known that the emission DA is sensitive to biases in the prior (e.g.: Tsikerdekis et al. 2022 ). I strongly recommend a calibration of the emission parameters prior to the assimilation experiments. Please give particular attention to the comments regarding the emission scheme in WRF by A. Ukhov (https://doi.org/10.5194/egusphere-2024-840-CC1)

The assimilated observations contain all aerosols, the simulated AOT and AE are calculated with all simulated aerosols, but only dust emissions are in the state vector. High AOD and low AE in Figure 5 indicate that the model is probably not simulating enough coarse particles (probably dust) over the sites. A simple calibration of the dust sources by a constant factor would be sufficient to improve the simulations in a similar way as the full data assimilation does here. This mismatch between the state variables and the observables is important, as it determines the ability of the data assimilation system to propagate the observational information into the model state, but it is only mentioned very superficially in the last line of the conclusions.   

 

The assimilated observations are from the Beijing area. This small spatial sample of assimilated observations, plus the way that the ensemble is constructed (by perturbing the emissions with a factor, one factor per ensemble), is likely to produce analyses emission corrections very homogeneous in the whole domain, meaning that the ratio between the analyses and first guess emissions is constant in the domain: increments of emissions in the western boundary (India and Pakistan) of Figures 1, 2 and 3 show this behaviour: the increments over Indian and Pakistan regions are similar to those of the Gobi desert, while the assimilated observations are located near Beijing. The temporal behaviour of the increments seems to show the same issue (Figure 4). 

The assimilation and verification period is extremely short: 4 days with very few observations (see for example Figure 16). It may be difficult to draw conclusions from such a small observational sample. A longer study period would be highly beneficial, although it is not always possible to perform.  

I haven't found strong evidence in the manuscript showing that the assimilation of AE improves the temporal and spatial variability of emissions or AOT. Also, it would be good to explain why AE assimilation increases AOD and consequently improves the skills. The scientific reason for this is not clear in this paper. It could be because it actually introduces important size information to the system, or just because it changes the balance between the observational and background terms in the cost function (adding more weight to the observations, for example), which in turn make the increments larger, which decreases the biases and improves the scores. 

I would suggest that, in order to make the study more scientifically attractive and with more enriching conclusions, the experiments should start from a relatively unbiased FR. Also, as noted by some of the co-authors in a previous article (Dai et al. 2019, last paragraph of their conclusions), the design of ensemble (with no temporal and spatial variability) and the state vector (only dust emissions) or other observation operator errors can negatively afffect the results. Given the strong constrain on the time and location of the assimilated observations (and their spatio-temporal representativity) these issues are, in this case, utterly important. 

Other comments:

Abstract : 

L15: Please note that AE does not resolve particle size. It is sensitive to it, but not only.

L17: I would suggest replacing the work "crucial" with "possible" or similar.

L21: The authors could replace the "official WRF-Chem" with the emission option they actually used.

L26: Is it necessary to distinguish between Mongolia and China Gobi?

Main text

L30: Please revise the first sentence. 

L43-44: It is very unlikely that we could fully understand the whole dust cycle by studying just one dust storm. Also, the logic of the argument and the connections between the sentences in these two lines are not clear. 

L45: Please delete "simulation".

L61: Please add "domain" or similar: "... forecast in the global domain".

L64-65: These lines are important for the manuscript, but not very precise. Please try to rewrite them.

L66 : Not so open. Please refer to the series of papers by Tsikerdekis et al. (ACP 2021, GMD 2022 and ACP 2023), where there is no explicit sensitivity of emissions to AE, but their usefulness can be inferred.

L67-69: The authors just said that it was an open question, and they answer it here by saying that ground-based AE is critical for dust emissions. I think the authors are willing to say that ground-based AE is better than satellite-based AE, and can be more useful for optimising dust emissions. Please note the accuracy of satellite retrieved AE depends on the instruments (not the same for MODIS than for multi-angle polarimeters), and on the retrievals algorithm. 

L70: The authors say here that they use the fixed-lag Kalman smoother (and later LETKF). Can it be provided a reference to the method that was actually used? Is it Dai et al. 2019?

L80: No need for "direct". Or  it is referring to the "direct sun" product from AERONET?

L84: This is inaccurate. It should be clearer with the formula or point to a good reference. 

L86: This 0.05 seems a little arbitrary. Is there any reference for the accuracy of AERONET 440-870 AE?

L90: Why 0.025? how do was estimated this value?. So far the reader does not know the spatial resolution of the WRF grid. 

L108: "... with 45 km ..."

L115: size bins in diameter or radius?

L117: MOSAIC acronym, description and need for use it is not clear here. Isn't this all done with GOCART? 

L128 So the ensemble only has variability in dust emissions? Does this mean that all AOD and AE departures (AERONET minus model), where all aerosols are included, are only attributed to increases in dust emissions? 

L131-132: This is not clear. Please clarify. 

L133: Please clarify if the authors are using the lagged LETKF (e.g. Schutgens 2012), the EnKS from Dai et al. 2019, or if it is the same method. 

L137: Please revise if this is the Kalman gain. The Kalman gain in this context is clearly defined in Eq. 10 of Hunt et al. 2007, where it is the matrix in the left multiplication of the departures (in observational space), not the vector w (in ensemble space). 

L152: Again, 4D-LETKF or EnKS? 

L165: Before showing the DA results, it would be useful to know exactly which and where observations were assimilated: for example, it would be sufficient to show their geographical location, time series and comparison with the FR performance in a qualitative sense. Also, if possible, the storm could be described in terms of AOD and AE from satellite observations, and if available, time-series of ground concentrations of TSP, PM20 or PM10 could useful. The later can be also useful as complement to the lidar verification in the results section.

L182-183: Interpretation of aerosol composition from a single value of AE does not seem a good practice. 

Figures : If not already implemented, please try to follow the ACP recommendations on colour scales, font size, etc. 

L193: It is not the best option to have different 3 figures for the 3 consecutive days.

L202-203 Again, a simple plot of observations and model AOT and AE will be useful at this point (Fig 5 could also be referenced). It may just be a model bias (e.g. of other aerosols) that makes the difference.  

L204-206: So far there is no information in the manuscript to support such a statement. This could be partly solved by showing the assimilated locations and time series beforehand. 

L213: The usual terminology indicates that it is the model or the observation operator that is inverted, not the emissions. The word "inverted" could be replaced with "posterior". 

L218-221: First presentation of the assimilated observations. This information needs to be presented before the results section. 

L223: Probably the word "obviously" is not really needed. 

L228-230: Evaluation scores are computed by mixing assimilated and non-assimilated observations rather than separately, making the evaluations difficult to interpret. 

L233-234: The authors claim:  "... and thus capture the aerosol spatiotemporal variation characteristics during dust transportation". True or not, this claim is not fully supported by the above results. Please revise this statement, as it is not obvious that the DA is doing more than just scaling the dust burden.

Figure 5: Please define this MFE. The standard definition is in the range [0,2].

L248-258: Differences between model experiments are clear, but it is extremely difficult to draw conclusions about model skill with such localised and small numbers of data points. 

L258: Again, AE could be useful to improve emissions, but it is not required.

L280 and 282: Please replace  AOD+AR by AOD+AE. Please follow the ACP guidelines for figure composition.  

L319: Please rewrite "in better consistent". 

L324: "Innovative superiority" may be an overstatement.

References :

Dai, Cheng, Goto, Schutgens, Kikuchi, Yoshida, et al. Inverting the East Asian Dust Emission Fluxes Using the Ensemble Kalman Smoother and Himawari-8 AODs: A Case Study with WRF-Chem v3.5.1. Atmosphere, 10(9), 543. https://doi.org/10.3390/atmos10090543, 2019.

Hunt, B., Kostelich, E. J., Szunyogh, I. Efficient data assimilation for spatiotemporal chaos: A local ensemble transform Kalman filter,B. Physica D: Nonlinear Phenomena, Volume 230, Issues 1–2, 2007, Pages 112-126, ISSN 0167-2789, https://doi.org/10.1016/j.physd.2006.11.008.

Schutgens, N., Nakata, M., and Nakajima, T. Estimating Aerosol Emissions by Assimilating Remote Sensing Observations into a Global Transport Model. Remote Sensing, 4(11), 3528–3543. https://doi.org/10.3390/rs4113528, 2012.

Tsikerdekis, A., Schutgens, N. A. J., and Hasekamp, O. P.: Assimilating aerosol optical properties related to size and absorption from POLDER/PARASOL with an ensemble data assimilation system, Atmos. Chem. Phys., 21, 2637–2674, https://doi.org/10.5194/acp-21-2637-2021, 2021. 

Tsikerdekis, A., Schutgens, N. A. J., Fu, G., and Hasekamp, O. P.: Estimating aerosol emission from SPEXone on the NASA PACE mission using an ensemble Kalman smoother: observing system simulation experiments (OSSEs), Geosci. Model Dev., 15, 3253–3279, https://doi.org/10.5194/gmd-15-3253-2022, 2022. 

Tsikerdekis, A., Hasekamp, O. P., Schutgens, N. A. J., and Zhong, Q.: Assimilation of POLDER observations to estimate aerosol emissions, Atmos. Chem. Phys., 23, 9495–9524, https://doi.org/10.5194/acp-23-9495-2023, 2023.