The main goal of this manuscript is to present a machine learning method to classify airglow image data. Narrow field-of-view images are separated into three classes: cloudy, clear but calm, clear with dynamic structures, in order to retrieve and analyse turbulent features associated to gravity wave breaking.

The machine learning method is very well explained, with a clear description of each parameters and steps. Results are also discussed in order to improve the technique. At the end, a small analysis of several dynamic structures is also presented.

The manuscript is well-written and very detailed, but the authors should address the following comments:

- Out of curiosity, is there a specific reason for pointing the imager toward azimuth 204 degrees?

- l. 95-98: hasn't fladfielding already removed reflection of the lens for example?

- Do you know if the star field has any effects on texture-based parameters like homogeneity or uniformity? Or on the FFT analysis?

- A table summarizing the 8 features would be helpful to remember what they are, and in which way they characterize the images.

- l. 268: How did you select the value -0.2? Does it come from another study or is it empirical?

- l. 309-316: The authors say that only the psd feature group shows some significant effect on the precision. Would it be worthy to remove the basic feature group, for example (faster calculations with still good results)?

- Tables 1 and 2 could be combined in only one table XXXX (0.YY)

- You realized that a lot of the mispredicted "calm" or "dynamic" images are in fact misclassified. How could you improve the classification?

- The mean average precision is 0.82, but only 0.63 for "dynamic" images (which is what you are looking for). It doesn't seem that good!

- The authors say that only using 8 features in the neural network is faster than the full images, but calculating these features requires already a lot of calculations (like for the FFT).

- Would it be possible to apply this method, or a similar method, on raw data? OF course, it would be a problem for the psd group! So, maybe not possible.

- Only 13 dynamic features in 8 months (~1.5 per month). How does it compare with previous studies from the same authors?

Minor edits:

l. 26: emission

l. 31: VIIRS

l. 43: ...turbulence, energy...

l. 52: ...assume that...

l. 59: remove . after "years"

l. 62: presents

l. 69: imbalance

l. 72: is presented in more detail

l. 80: remove parenthesis (already in l. 74)

l. 90: database

l. 91: measurements... during 258 nights.

l. 94: remove "as soon as it is"

l. 92: protecting window

l. 122: replace "dynamic" by "cloudy"

l. 129: ...high if the image...

l. 140: remove "the fact"

l. 150: From the...

l. 155: , respectively.

l. 237: a neuron

l. 315: ...measures is the psd...

l. 321: This configuration was modified according to the validation data...

l. 336: and Table 3

l. 365: and Table 2

l. 393: In the next step...

l. 419: incorrect or correct?

l. 430: "calm"

l. 440: ...means that this is...

l. 492: database

l. 499: ..."calm" airglow, or being...

l. 504: remove "All in all"