This manuscript presents a neural network model in order to detect volcanic ash clouds using Sentinel-3 SLSTR (Sea and Land Surface Temperature Radiometer) daytime products. The neural network is trained with MODIS daytime imagery from the Eyjafjallajökull eruption in May 2010. Then it is applied to the Raikoke eruption in June 2019. The results show that the neural network model can accurately detect volcanic ash from Raikoke compared with RGB visual inspection and BTD (Brightness Temperature Difference) procedure. Moreover, the plumes identified by neural network model agree well with the plume manually identified for the specific SLSTR images.

The manuscript is very well structured and written. It addresses an important issue in detection of the volcanic ash clouds and presents a solution which is beneficial for remote sensing and modeling volcanic ash dispersion. The methods and assumptions are scientifically sound and the results are well elaborated. Thus, I recommend the manuscript for publication after adressing the following points:

1- The authors should use/cite the published data instead of relying on private communication (L92). Specifically, there are several papers in this special issue that present ash and SO2 mass (e.g. Muser et al 2020, ACP). I strongly suggest that the authors review the published papers related to Raikoke and use them in the introduction and discussions. 

2- Raikoke and Eyjafjallajökull are both high-latitude volcanoes. How would the model perform on tropical eruptions like la Soufrière 2021? Is the model transferable to tropical conditions or different ash compositions? It will be interesting to see the application to la Soufrière.

3- I would like to see the R² and RMSE of the neural networks during training, validation and test. The topology of the neural network model (large number of neurons in the hidden layer) and split of the training/validation/test might lead to overfitting. Besides, please add info about the training method.   

4- There are no discussions on the uncertainty and the limitations of the presented model. 

Specific comments: 

L32-34: this part is not precise. Ash is a part of tephra with D<2 mm. Then we have fine and very fine ash. Please revise.

L49: you mean ΔT_{11μm - 12μm} ?

L70-72: NNs are good for what they are trained for. Their transferability to other eruption at different altitudes and with different ash composition (optics) might be challenging. Please comment on this.   

L159: What is the measure of accuracy? R²?

L205: for consistency, use "meteorological clouds" in the whole manuscript. 

L226: this argument is too strong. See my previous comments. 

Tables 4 and 5: It is very difficult to make any quantitative conclusion from these tables. Use other quantitative measures like SAL. 

The study ``Volcanic cloud detection using Sentinel-3 satellite data by means of neural networks: the Raikoke 2019 eruption test case'' by Petracca et al. introduces a scene classification algorithm for the Sentinel-3 Sea and Land Surface Radiometer data based on neural networks. The classification is applied in a case study of the eruption of the Raikoke volcano in 2019. While the focus is on detecting volcanic ash plumes the classification mask also provides information on the surface, underlying surface under volcanic ash, and clouds. Although the paper is well structured and written I miss substantial information on the neural network. No information on how it was coded nor the source were provided. Moreover the results presented in this study lack a comparison with already published findings on the Raikoke eruption and measurements by other instruments. Hence I'd recommend a major revision before publication. 

General comments:

In the introduction solely volcanic ash measurements in the mid-infrared are discussed. However the SLSTR mainly has channels in the VIS to near infrared spectral range. I suggest to also introduce VIS/near-IR volcanic ash measurements.

Throughout the manuscript ``weather clouds'' are mentioned. Please specify what you mean. Ice clouds, liquid clouds, mixed phase clouds, or all?

The description of the case study on the Raikoke eruption lacks references. Please have a look at the publications in this special issue to verify your reconstruction of the plume (in Fig. 1) and to substantiate your estimates of SO2 and ash.

The methodology section I found somewhat confusing. Maybe separate the instrument description from the method description. The description of both instruments, MODIS and SLSTR, lack some information. What is their spectral range? What is their equatorial crossing time? Since when are they operating? What is the oblique view of SLSTR, which is mentioned later? Which data products were used? First I had the impression that the classification categories (Ash over sea, ash over clouds, sea surface, ...) are MODIS products. Only later I realized that you made up these categories manually from MODIS Eyjafjallajökull observations. Please improve the description.

Concerning the neural network, how did you build the network? Did you use Python and some packages? Did you use anything else? Please provide more information. Also you mention the time benefit of using NNs. How much time did it take to train the NN? How long does it take to analyse a scene with the NN compared to the BTD method? When mentioning the speed advantage, please provide numbers/measurements.

When comparing the results from the BTD-method with the results of the NN-approach, please comment on the sensitivity of both methods (BTD and NN), as well as the manual detection in the VIS, on the ash AOD. Why should the BTD-approach lead to false positives in the case of the Raikoke?

I clearly disagree that Section 4.1 is a validation of the method. The reference is tuned towards an ash plume discernible in RGB satellite images. The detection sensitivity towards ash/aerosol AOD in nadir geometry and VIS spectral range is different to other wavelengths and satellite measurement geometries. Since the NN method relies on multiple wavelengths ranging from VIS to mid-IR, the results should be compared to VIS to mid-IR standard ash/aerosol detection products. Why don't you compare with measurements of other instruments, e.g. TROPOMI, AIRS, IASI, OMI, GOME-2, CALIPSO?

Moreover, the application of this method to only 2 scenes of a single volcanic eruption, measured on the same day is rather inconclusive. Please consider applying the NN method to other volcanic eruptions (as e.g. Gray and Bennartz, 2015, tested their NN approach to 7 volcanic eruptions). Also, how would you method deal with desert dust, which is a challenge to the BTD approach?

Specific comments:

l33-34: Please specify coarse and fine in µm.

l34-35: Volcanic plumes also have a liquid part, as formation of sulfate aerosol starts immediately e.g. see Glasow et al. (2009).

l60: When mentioning other volcanic ash detection algorithms, please also consider Gangale et al. (2010) and Clarisse et al. (2013).

l68: I wonder why you are referring to two studies using NNs for ozone retrievals, although sufficient examples for aerosol and clouds are already mentioned.

l87: What does near the vent mean? Please specify the radius around the volcano from which the BT of the plume was derived. Also what does ``some distance upwind'' mean? Was it always the same distance? Which criteria did you apply?

l92-94: Please remove speculations about the water vapour. 

l100: Please explain what is a ``multilayer perceptron neural network''?

l108: What is the difference between Sentinel-3A and 3B?

l109: Which procedure is meant here? I don't understand why this is mentioned after the instrument description.

Table 1: Please provide consistently the bandwidth for both instruments. Did you use all channels in the NN?

Fig. 2:  Do the text ``Neural Network'' and the picture mean the same, or are this two different neural networks?  Also there are two arrows from SLSTR to both? networks leading to one classification. Are two different networks used for the classification?

l116: What does ``nine MODIS data'' mean? Is it 9 days of data? Is it 9 swathes? Is it 9 images? Please indicate the lat/long region around Eyjafjallajökull that was selected.

l117: What does pattern mean? Is pattern=pixel?

l133-141: Where and how large are the uncertainties of your ground truth? Are you considering the visual classification of RGB-images as the reference?

l153-154: Is the a posteriori filter only applied to the categories ``land'' and ``sea'' or also to ``ash over land'' and ``ash over sea''?

Fig. 5: What are the red and cyan color in the RGB image? Was the ``Not classified'' class only applied to ``Sea'' and ``Land'', or also to ``Ash_sea'' and ``Ash_land''?

l181: Does ``... difference between ... channels S8 and S9...'' mean mean radiance (S8) - mean radiance (S9)?

Fig. 6: Fig. 6a shows many contrails, but in Fig. 6c only few of them are classified as ``Cloud_ice''. Can you comment on this? Why are so many classified as ``cloud'' that was introduced as liquid cloud and which rather represents low altitude clouds?

l189: What do you mean by "pixels identified as volcanic cloud but that are not below the volcanic cloud..."? Please clarify.

l198-199: Here  you state, that some pixels were misclassified as ``ash_land'' instead of ``ask_sea''. But shouldn't it rather be ``ash_cloud''? Most of the area around Raikoke is marked as ``cloud'' or ``ice cloud''. It would be surprising if only the region below the volcanic ash plume is not covered by clouds.

l206: What do you mean by ``water vapour cloud''? In the RGB images only ice, liquid water, or mixed clouds are visible.

l208: Having VIS RGB images at midnight sounds strange. I assume you mean 0 UTC.

201-214: Why do you think the BTD approach produces wrong positive results in the case of the Raikoke eruption (Fig. 6)? Please explain. I'd rather consider the BTD ash plume realistic, because it pretty much resembles the SO₂ plume shape measured by TROPOMI on 23 June (e.g. Leeuw et al., 2021, Cai et al., 2022). How do you know that there wasn't any ash above the contrails and these underlying clouds enhanced the ash signal of the otherwise ``thin'' ash layer, which remained invisible in regions without underlying cold clouds (=high altitude clouds)? 

l220: Do you mean higher opacity here?

Fig. 7c,d: Why are mostly clear regions (43-33N, 170-175E) classified as ``Cloud''? Please comment.

l224-226: What do you mean by ``different scenario''? In terms of season, latitude, and injection height, the training eruption is similar to the showcase of the Raikoke eruption.

Fig. 9: What does the white color indicate? Why does the CSCM detect clouds in apparently clear regions?

l274: Again, what are ``meteo clouds'' and ``meteo ice clouds''? Liquid and ice clouds?

Technical comments:

l26-27: remove ``it'' -> ...which is...

l27: manually -> manual

l30: NN, please introduce abbreviations

l33: by -> of

l49: region -> regions (2x)

l66: in -> at

l84: AHI, please introduce abbreviations

Fig1 caption: was -> were; does -> do

l198: ash-on-land -> ash-over-land

l212: respect -> with respect

References:

Roland von Glasow, Nicole Bobrowski, Christoph Kern: The effects of volcanic eruptions on atmospheric chemistry, Chemical Geology, Volume 263, Issues 1–4, 2009, Pages 131-142, https://doi.org/10.1016/j.chemgeo.2008.08.020

G. Gangale, A.J. Prata, L. Clarisse: The infrared spectral signature of volcanic ash determined from high-spectral resolution satellite measurements, Remote Sensing of Environment, Volume 114, Issue 2, 2010, Pages 414-425, https://doi.org/10.1016/j.rse.2009.09.007

Clarisse, L., Coheur, P.-F., Prata, F., Hadji-Lazaro, J., Hurtmans, D., and Clerbaux, C.: A unified approach to infrared aerosol remote sensing and type specification, Atmos. Chem. Phys., 13, 2195–2221, https://doi.org/10.5194/acp-13-2195-2013, 2013.

Gray, T. M. and Bennartz, R.: Automatic volcanic ash detection from MODIS observations using a back-propagation neural network, Atmos. Meas. Tech., 8, 5089–5097, https://doi.org/10.5194/amt-8-5089-2015, 2015. 

de Leeuw, J., Schmidt, A., Witham, C. S., Theys, N., Taylor, I. A., Grainger, R. G., Pope, R. J., Haywood, J., Osborne, M., and Kristiansen, N. I.: The 2019 Raikoke volcanic eruption – Part 1: Dispersion model simulations and satellite retrievals of volcanic sulfur dioxide, Atmos. Chem. Phys., 21, 10851–10879, https://doi.org/10.5194/acp-21-10851-2021, 2021. 

Cai, Z., Griessbach, S., and Hoffmann, L.: Improved estimation of volcanic SO2 injections from satellite retrievals and Lagrangian transport simulations: the 2019 Raikoke eruption, Atmos. Chem. Phys., 22, 6787–6809, https://doi.org/10.5194/acp-22-6787-2022, 2022. 

The Authors present a neural network technique to detect volcanic ash clouds

by combining visible and thermal infrared channels of moderate resolution spectroradiometers.

A neural network trained on MODIS imagery acquired during the Eyjafjallajökull eruption in 2010

is applied to two pairs (nadir and oblique view) of SLSTR images of the Raikoke eruption that occurred in 2019.

The neural network method is compared to the classic brightness temperature difference (BTD) method and the

accuracy of the two methods is evaluated against manually classified pixels. The results show a reasonable performance

of the NN method in detecting ash clouds in nadir view, whereas I have some perplexities about its performance in the oblique

view, as it seems to me that the NN underperforms for a fairly thick ash plume which I would expect to be easily detectable.

All in all, I think the paper can be published with minor revisions, although I recommend careful proofreading by a native

English speaker, as the quality of the written English does not look impeccable to me.

MAIN COMMENT

I think that the extension of the NN approach to oblique view needs further investigation.

What are the typical values of the viewing angles sampled in the oblique view and how do they

compare to those of the nadir view? If the air mass sampled in the oblique view is much bigger

than that sampled in nadir view, the difference in the slant optical depth may translate to

a npticeable difference in top-of-atmosphere signal levels. Furthermore, if there is a large difference

in the observed scattering angles you may be also sampling different ranges of (weather and ash)

cloud phase functions, which also may lead to significant differences in the signal levels in VIS/NIR channels.

In this case, it looks far from obvious to me that the NN can still be applied reliably to oblique view

situations that are probably not covered in the training set.

Therefore, I would recommend studying the sensitivity of the NN detection to the observation angle

by generating synthetic top-of-atmosphere spectra of VIS/NIR radiance and thermal brightness temperature

for a typical liquid water, ice and ash cloud. In my opinion, the results presented in the paper 

do not allow to draw reliable conclusions on the robustness of the NN method to off-nadir observations.

DETAILED COMMENTS

- P1, L15, spaceborne sensors acquired data -> satellite data

- P1, L19, The classification of the clouds and of the other surfaces -> A classifications of clouds and other surfaces

- P1, L22, foster the robustness of the approach, which allows overcoming -> allow to extend the approach to SLSTR, thereby overcoming

- P2, L42, to detect the volcanic cloud -> to detect volcanic clouds

- P2, L43, you can remove "problem" after detection

- P2, L43, lies on -> relies on

- P2, L44. There is no such thing as "water vapour clouds". I guess you mean "liquid water clouds"

- P2, L49, region -> regions

- P2, L59, procedures described -> described procedures. Plus, is "among" really what you mean, or do you mean "in addition to"?

          Does "described" refer to Prata et al. (2001b) and Corradini et al. (2008,2009)?

- P3, L70, statistical -> statistically

- P3, L71. Is this real time capability really an advantage of the NN approach? Isn't the BTD method also in near real time,

          given that it involves taking a difference? Furthermore, in an emergency scenario is there really such a big advantage in

          correctly detecting a few more ash pixels than the BTD method? 

          

- P3, L85-86, either "a vertically ascending cloud" or "vertically ascending clouds"

- P4, L103, "water vapour" -> "liquid water"

- P4, L107. At what angles does SLTR dual view observe?

- P4, L109-110. I don't understand the use of "since" here. What do you mean when you say that the  feasibility of the method was confirmed for high latitudes "since" your study area  is at medium-high latitudes. 

            

-P8, L159-160. How is each percentage in the confusion matrix computed? Furthermore, overall accuracy is not a particularly informative parameter.  Given that the main focus is on ash, it may be useful to provide statistics on the task of ash detection  (probability of detection, false alarm ratio, critical success index).

                   

-P8, L166. What do you mean by "commission" and "omission" errors? I guess one is "false detection" and the other is "missed detection", 

           but it is not clear which is which.      

           

-P9, Figure 5. If I look at panel a, it seems to me that on the edges of the plume there are quite a few pixels that the NN classifies

         as "cloud ice". Do you have an idea why this happens?

         

-P11, L186, emphasizes -> shows

-P11, L198-199, "even if some pixels are misclassified as ash on land". For such a thick ash cloud I would indeed expect that there is hardly any information in the signal to distinguish ash over land from ash over sea or cloud. Does it really make sense to introduce such a fine distinction 

 between ash classes? What do you gain from that?

           

-P11, L199, less false positives -> fewer false positives. On top of that, are they really false positives? Doesn't the BTD detect fewer ash pixels

 compared to the NN?

         

-P11, L206, water vapour cloud -> liquid water cloud

-P11, L213, aerial trails -> aircraft condensation trails

-P11, L214. What causes the BTD method to give false positives over contrails?

-P11, L218, "produces good results". I would say "reasonable". The ash cloud looks so thin here that  I doubt you have a very good reference to compare your results against. How does the BTD approach perform for this image?

-P12, Fig. 7(a). Here the NN seems to detect a much smaller portion of the plume compared to what happens in the nadir image.

      Interestingly, a large fraction of the ash cloud is again classified as cloud ice (see my previous comment about Fig. 5).

      Do you have any explanation for this apparently systematic tendency to confuse ash with cloud ice? Again, how does the BTD

      approach perform for this case?      

        

-P12, L220, "this is due to the opacity of the volcanic cloud". Why is the opacity of the volcanic cloud a problem for the NN detection? 

         I would expect a more opaque cloud to provide a better contrast against weather clouds. 

         Even visually, the plume in image (a) looks easier to detect than the faint plume in image (c).

           

-P13, L249, matching -> agreement

- P14, Table 4. How would such a table compare to a similar one for BTD vs MPM?

- P16, L286, includes also -> also includes