• It would be useful to add information to Fig. 6 to ease comparison with the Bony et al 2020 Iorg/S quadrants. In Line 150, you label the quadrants (e.g. Fish in the upper right corner), and perhaps such a labeled grid could be added to Fig. 6.
• I also find the symbols a bit small and sometimes hard to distinguish the colors/patterns. The figure is, however, striking regarding the good agreement between the manual classification and the deep neural network.
• Could you elaborate briefly on the lack of patterns identified in the bottom right corner? e.g. why flowers are more centered (line 150)?

• Can you explain why the deep neural net seems to identify a higher area fraction covered by each pattern? It also seems like there is a higher area fraction covered by each pattern for the IR (center bar) than the visible (left bar)

Schulz presents an in-depth and well-written description of a very useful

dataset of manual classifications of shallow cumulus clouds covering the

EUREC4A campaign period. The comparision with more traditional techniques for

measuring organisation provides a valuable reference for interpreting the

manual classifications. And contrasting the results of using different

datasources is very valuable. The detail with which the dataset is described

and the openeness by which the tools used have been shared is commendable and

inspiration for our community as a whole.

There are two aspects I would find very valuable to giving a little more depth.

First, is the definition of "truth" in this manual classifications. This would

focus on answering questions such as "which of workflow dataset should we trust

as the most truthful and why?" "is it possible to produce a kind of consensus

amond the four workflows?" This could draw in prior studies using manual

classifications referenced in this publication. Second, currently there is

little analysis of the manual classifications created for the simulation data

(as compared to the obsevation-based workflows), why is this and was this

analysis done?

Detailed comments

- [p 1 l 21] "studies concentrated on the classification of meso-scale patterns...":

  - You could also mention Denby 2020 here as an example of using a neural

    network for classification. It would also be good to mention Wood &

    Hartmann 2006 here already. You should also include reference to Janssens

    et al 2021

    (https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2020GL091001) for

    it’s review of traditional metrics

- [p 3 l 59] "as it quickly turned out that the identification of the patterns in

  the model simulation was too demanding. The features had too little

  similarity with those found in nature.":

  - I think this needs reformulating. “Accumulated intentionally” isn’t so

    clear, you mean that people tended to label the observations and not the

    simulation output? You could refer to the totals in Figure 2 to make this

    point. I would also emphasise that you are continuing to talk about

    classifications made in the ICON workflow in the following sentences.

    Because the you lead the paragraph with the total number of classifications

    I initially read this as if “sugar” was hard for people to identify across

    all workflows (but that isn’t the case I think, cf Figure 2)

- [p 5 l 95] "This process eliminates overlaps of same-user classifications for

  each pattern and turns the data into masks, rather than coordinates (see Fig.

  3)":

   - What happens if the user draws two bounding-boxes with different types of

     classification that spatially overlap? (It shouldn’t really happen, but

     maybe you could just state that it didn’t)

- [p 6] "Figure 4":

   - There is a bit of aliasing for the text in figure 4. Maybe storing as a

     pdf/svg would be better?

- [p 8 l 125] "Comparison with other classifications":

   - There doesn’t seem to be any analysis of the ICON simulation

     classifications. It would be good to mention why this was left out.

- [p 8 l 132] "org, Tompkins Adrian M. and Semie Addisu G. (2017)) with the

  mean cluster size (S)":

   - In choice of these metrics maybe you could mention how this fits with the

     findings for Janssens et al 2021?

- [p 8 l 136] "we focus on a domain size of 10×10 degrees to do the comparison":

   - I would rephrase to be clearer to something like “we calculate this

     metrics over a 10x10 degree sub-domain”. I would say “Specifically” rather

     than “Precisely” next. Also 10N to 15N is only 5 degrees, should it be 5N

     to 15N?

- [p 8 l 134] "detect the patterns in geostationary infrared images of 135

  GOES-16 ABI (Schulz et al., 2021)":

   - Does this mean that the network was trained on the dataset from the

     EUREC4A IR (channel 13) workflow to predict the manual classifications

     (masks) created by participants in the same workflow? It would be helpful

     to emphasise this, and specifically in say the caption of Figure 7, to so

     that the neural network was trained on the IR data (which would also

     explain why the agreement is better with the IR rather than the visible

     manual classification)

- [p 8 l 144] "we expect Gravel and Flowers to be rather regularly distributed and

  therefore to 145 have a lower Iorg compared to Fish and Sugar":

   - I don’t quite understand why “flowers” should be more “regularly

     distributed” than “sugar”? My understanding is that “sugar” is scatter

     small cumuli which I would expect to be very regularly distributed.

- [p 8, Figure 6] "mean cluster size (S)":

   - What are the units of “mean cluster size”? Are they pixels?

- [p 9 l 150] "we applied a threshold of 0.1 on the frequencies":

   - This is a bit unclear. Above you talk about the “percentage of agreement”

     but here of “frequencies of the level 3 dataset”. How do you go from

     “percentage of agreement” to “frequency”? Are they the same? Does a

     threshold of 0.1 mean that only 10% of participants needed to say that

     they label an area as a given pattern? Is that a reasonable number? It

     seems to quite low to me, wouldn’t that lead to very large masks? What

     happens if two users classify a given pixel with two different labels (I

     don’t think that is taken into account with the current calculation, but

     the word “frequency” suggests to me that it should)?

   - Also, having a cumulative area fraction larger that 1.0 is quite

     confusing. It would be good to discuss what that means and why it is

     reasonable. How would I read from Figure 7 what area fraction is

     unclassified?

- [p 9 l 162] "While the Iorg/S metric is computationally cheap and can be easily

  applied to different regions, the manual classifications are naturally more

  accurate":

   - This doesn’t quite follow for me. Figure 7 isn’t attempting to produce

     classification into the four organisation patterns using only Iorg/S, and

     so I don’t think this analysis shows that it isn’t possible with just

     Iorg/S.

- [p 9 l 164] "manual classifications are most 165 accurate":

   - I would like to understand a little better how you draw this conclusion.

     What is your measure of accuracy? Aren’t the manual classifications being

     used as “truth” here? If so, and assuming that any other method will

     classify differently in some way, how can any other method predict

     something better than what is being used as the reference?

- [p 10 l 179] "demands further investigations on how":

   - should be "demands further investions into how..."